import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import Dataset
from torch.utils.data import DataLoader

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

from tqdm import tqdm
from time import sleep
import os
from os import listdir
import matplotlib.image as mpimg
from matplotlib.animation import FuncAnimation as FA
import random

pd.set_option('display.max_columns', 50)
pd.set_option('display.max_rows', 50)

# Set Parameters for Training
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'  # Use serif font for better readability
mpl.rcParams['font.size'] = 12  # Consistent font size
mpl.rcParams['axes.linewidth'] = 1  # Thicker axes lines
mpl.rcParams['xtick.major.size'] = 5  # Ticks adjustments
mpl.rcParams['ytick.major.size'] = 5
mpl.rcParams['xtick.major.pad'] = 5
mpl.rcParams['ytick.major.pad'] = 5
mpl.rcParams['axes.labelsize'] = 14
mpl.rcParams['axes.titlesize'] = 16
mpl.rcParams['legend.fontsize'] = 12
mpl.rcParams['figure.dpi'] = 300  # High DPI for publication quality
mpl.rcParams['savefig.bbox'] = 'tight'  # Tight bounding box
mpl.rcParams['axes.grid'] = True  # Enable grid by default
mpl.rcParams['grid.alpha'] = 0.5 # Grid alpha
mpl.rcParams['lines.linewidth'] = 1.5 # Line width
mpl.rcParams['scatter.marker'] = 'o' # Marker

# import time

# # Generate a random seed based on the current time
# seed = int(time.time())
# print("Selected seed:", seed)

seed = 42
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)

folder_path = './CMAPSSData/'

listdir(folder_path)

file_name = 'FD003.txt'

df_train = pd.read_csv(folder_path + 'train_' + file_name, header = None, sep = ' ')
df_test = pd.read_csv(folder_path + 'test_'+file_name, header = None, sep = ' ')
rul_test = pd.read_csv(folder_path + 'RUL_'+file_name, header = None)

for df in [df_train, df_test, rul_test]:
    display(df.head())

col_names = []

col_names.append('unit')
col_names.append('time')

for i in range(1,4):
    col_names.append('os'+str(i))
for i in range(1,22):
    col_names.append('s'+str(i))

df_train = df_train.iloc[:,:-2].copy()
df_train.columns = col_names
display(df_train.head())

df_test = df_test.iloc[:,:-2].copy()
df_test.columns = col_names
display(df_test.head())

units_training = max(df_train['unit'])
units_testing = max(df_test['unit'])
print(f'Units in the training dataset: {units_training}')
print(f'Units in the testing dataset: {units_testing}')
df_train = df_train[df_train['unit'] != 260]
if units_training > units_testing:
    df_train = df_train[df_train['unit'] <= units_testing]
elif units_training < units_testing:
    df_test = df_test[df_test['unit']  <= units_training]

assert max(df_train['unit']) == max(df_test['unit'])
no_units = max(df_train['unit'])
print(f'No of units in training and testing dataset after: {no_units}')

Units in the training dataset: 100
Units in the testing dataset: 100
No of units in training and testing dataset after: 100

MAX_RUL = 125
no_units = min(rul_test.shape[0],max(df_train['unit']))
print(f'units :{no_units}')

units :100

rul_list = []
engine_numbers = no_units
for n in np.arange(1,engine_numbers+1):
    
    time_list = np.array(df_train[df_train['unit'] == n]['time'])
    length = len(time_list)
    rul = list(length - time_list)
    rul = [min(MAX_RUL,x) for x in rul]
    rul_list += rul
    
df_train['rul'] = rul_list

rul_list = []

for n in np.arange(1,engine_numbers+1):
    
    time_list = np.array(df_test[df_test['unit'] == n]['time'])
    length = len(time_list)
    rul_val = rul_test.iloc[n-1].item()
    rul = list(length - time_list + rul_val)
    rul = [min(MAX_RUL,x) for x in rul]
    rul_list += rul

df_test['rul'] = rul_list

for df in [df_train, df_test]:
    display(df.head())

from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler

# Initialize KMeans and StandardScaler
kmeans = KMeans(n_clusters=6, random_state=42)
scaler = StandardScaler()

# Apply clustering for specific datasets
if file_name in ['FD002.txt', 'FD004.txt']:
    # Ensure required columns are present
    if {'os1', 'os2', 'os3'}.issubset(df_train.columns):
        # Select operating conditions and scale them
        operating_conditions = df_train[['os1', 'os2', 'os3']]
        scaled_conditions = scaler.fit_transform(operating_conditions)

        # Apply k-means clustering and assign operation_mode
        df_train['operation_mode'] = kmeans.fit_predict(scaled_conditions)
    else:
        raise ValueError("Columns 'os1', 'os2', and 'os3' are missing in the dataset!")
else:
    print(f"No clustering applied for dataset {file_name}.")

if file_name in ['FD002.txt', 'FD004.txt']:
    df_test['operation_mode'] = kmeans.predict(scaler.transform(df_test[['os1', 'os2', 'os3']]))

No clustering applied for dataset FD003.txt.

for df in [df_train, df_test]:
    display(df.head())

sensor_colums = ['s'+str(i) for i in range(1,22)]
print(sensor_colums)

['s1', 's2', 's3', 's4', 's5', 's6', 's7', 's8', 's9', 's10', 's11', 's12', 's13', 's14', 's15', 's16', 's17', 's18', 's19', 's20', 's21']

##  Z-score Normalization
if file_name in ['FD001.txt', 'FD003.txt']:
    # Dictionary to store mean and std for columns containing 's'
    mean_std_dict = {}

    # Calculate mean and std for 's' columns in df_train
    for c in sensor_colums:
        mean_std_dict[c + '_mean'] = df_train[c].mean()
        mean_std_dict[c + '_std'] = df_train[c].std()

    # Apply standard normalization to df_train
    for c in sensor_colums:
            df_train[c] = (df_train[c] - mean_std_dict[c + '_mean']) / (mean_std_dict[c + '_std'] + 1e-6)

    # Apply standard normalization to df_test using df_train's mean and std
    for c in sensor_colums:
            df_test[c] = (df_test[c] - mean_std_dict[c + '_mean']) / (mean_std_dict[c + '_std'] + 1e-6)

if file_name in ['FD002.txt', 'FD004.txt']:
    # Function to normalize each group based on its operation_mode
    def z_score_normalize_by_mode(group):
        group = group.copy()  # To avoid modifying the original data
        for c in sensor_colums:
            group[c] = (group[c] - group[c].mean()) / (group[c].std() + 1e-6)
        return group

    # Normalize training data
    df_train_normalized = df_train.groupby('operation_mode', group_keys=False).apply(z_score_normalize_by_mode)

    # Normalize test data using train stats for each operation_mode
    for mode in df_train['operation_mode'].unique():
        train_group = df_train[df_train['operation_mode'] == mode]
        test_group = df_test[df_test['operation_mode'] == mode]
        
        for c in sensor_colums:
            # Ensure the column dtype is compatible
            df_test[c] = df_test[c].astype(float)
                
            mean = train_group[c].mean()
            std = train_group[c].std()
                
            df_test.loc[df_test['operation_mode'] == mode, c] = (
                df_test.loc[df_test['operation_mode'] == mode, c] - mean
            ) / (std+1e-6)

    df_train = df_train_normalized.copy()

if file_name in ['FD002.txt', 'FD004.txt']:
    df_train = df_train.drop('operation_mode', axis = 1)
    df_test = df_test.drop('operation_mode', axis = 1)

os_colums = ['os'+str(i) for i in range(1,4)]
print(os_colums)

mean_std_dict = {}
# Calculate mean and std for 's' columns in df_train
for c in os_colums:
    mean_std_dict[c + '_mean'] = df_train[c].mean()
    mean_std_dict[c + '_std'] = df_train[c].std()

# Apply standard normalization to df_train
for c in os_colums:
    df_train[c] = (df_train[c] - mean_std_dict[c + '_mean']) / (mean_std_dict[c + '_std'] + 1e-6)

# Apply standard normalization to df_test using df_train's mean and std
for c in os_colums:
    df_test[c] = (df_test[c] - mean_std_dict[c + '_mean']) / (mean_std_dict[c + '_std'] + 1e-6)

# Display the first few rows of both datasets
for df in [df_train, df_test]:
    display(df.head())

['os1', 'os2', 'os3']

sample = 4
sample_df = df_train[df_train['unit'] == sample].copy()

# Select only sensor columns
n_sensors = len(sensor_colums)
rows = (n_sensors + 2) // 3  # Calculate rows for the grid

# Create subplots
fig, axes = plt.subplots(rows, 3, figsize=(15, 5 * rows))

# Flatten axes for easier iteration
axes = axes.flatten()

# Plot each sensor's data
for j, c in enumerate(sensor_colums):
    axes[j].plot(sample_df['time'], sample_df[c], label='Sensor data')
    axes[j].set_title(c)
    # axes[j].set_xlabel('Time')
    # axes[j].set_ylabel('Value')
    axes[j].legend()

# Hide unused subplots
for ax in axes[n_sensors:]:
    ax.axis('off')

# Adjust layout for better spacing
fig.tight_layout(pad=3.0)
plt.savefig(f'figures\\sensor_of_engine_{sample}_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()

sensor = 's2'
for sample in range(1,6):
    sample_df = df_train[df_train['unit'] == sample].copy()
    sensordata = sample_df[sensor].to_numpy()
    plt.plot(sensordata, label = "engine "+str(sample))
plt.grid()
plt.legend()
plt.ylabel('Sensor 2')
plt.xlabel('Cycle')
plt.savefig(f'figures\\sensor_{sensor}_of_engines_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()

import seaborn as sns
import matplotlib.pyplot as plt

# Compute correlation matrix
corr_matrix = df_train.corr()

# Plot heatmap
plt.figure(figsize=(20, 14))
sns.heatmap(corr_matrix, annot=True, cmap="coolwarm", fmt=".2f")
plt.title("Feature Correlation Matrix")
plt.savefig(f'figures\\Feature_Correlation_Matrix_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()

from sklearn.ensemble import RandomForestRegressor
import pandas as pd

Xt = df_train.iloc[:,2:-1]
yt = df_train.iloc[:,-1]

rf = RandomForestRegressor()
rf.fit(Xt, yt)

# Get feature importance
importance = pd.Series(rf.feature_importances_, index=Xt.columns)
importance = importance.sort_values(ascending=False)

print("Feature Importances:\n", importance)

Feature Importances:
 s11    0.632116
s9     0.140358
s14    0.031786
s4     0.025796
s12    0.024264
s7     0.020847
s13    0.018195
s8     0.015298
s15    0.013585
s3     0.012977
s21    0.012731
s2     0.012119
s6     0.010617
s20    0.009913
os1    0.009259
os2    0.005749
s17    0.004232
s10    0.000159
s5     0.000000
s1     0.000000
s16    0.000000
s18    0.000000
s19    0.000000
os3    0.000000
dtype: float64

# Combine thresholding and sorting
threshold = 1e-3
low_importance_features = importance[importance < threshold].sort_values()
print("Low importance features sorted:\n", low_importance_features)

Low importance features sorted:
 s5     0.000000
s1     0.000000
s16    0.000000
s18    0.000000
s19    0.000000
os3    0.000000
s10    0.000159
dtype: float64

df_train[low_importance_features.index].describe()

#Drop os3, s1, s5, s6, s10, s16, s18, s19 from both train and test
drop_cols1 = low_importance_features.index

df_train = df_train.drop(drop_cols1, axis = 1)
df_test = df_test.drop(drop_cols1, axis = 1)

for df in [df_train, df_test]:
    display(df.head())

ratio = 0.8                     #Ratio of training and validation datasets

units = np.arange(1, no_units+1)
no_selected = round(ratio * engine_numbers)
train_units = list(np.random.choice(units,no_selected, replace = False))
val_units = list(set(units) - set(train_units))
print(val_units)


train_data = df_train[df_train['unit'].isin(train_units)].copy()
val_data = df_train[df_train['unit'].isin(val_units)].copy()

[14, 15, 18, 24, 34, 40, 44, 48, 53, 62, 63, 72, 74, 78, 80, 82, 85, 87, 90, 94]

#Smoothing Function: Exponentially Weighted Averages

def smooth(s, b = 0.98):

    v = np.zeros(len(s)+1) #v_0 is already 0.
    bc = np.zeros(len(s)+1)  

    for i in range(1, len(v)): #v_t = 0.95
        v[i] = (b * v[i-1] + (1-b) * s[i-1]) 
        bc[i] = 1 - b**i

    sm = v[1:] / bc[1:]
    
    return sm

# s = [1,2,3,4,5]
# print(s)
# print(f'After Smoothing: {smooth(s)}')

#Smoothing each time series for each engine in both training and test sets
beta = 0.98
# if file_name=='FD002' or file_name=='FD004':
#     beta = 0.8

# Verify data integrity
assert 'unit' in df_train.columns, "The 'unit' column is missing in df_train"
assert 'unit' in df_test.columns, "The 'unit' column is missing in df_test"

def smooth_series(df, beta, unit_col='unit', sensor_prefix='s'):
    """
    Smooths time series for each engine and sensor column in the dataset.
    """
    for col in df.columns:
        if sensor_prefix in col:
            sm_list = []
            for unit in df[unit_col].unique():
                # Get sensor data for the current unit
                unit_data = df[df[unit_col] == unit]
                s = np.array(unit_data[col].copy())
                sm = list(smooth(s, beta))  # Apply smoothing
                sm_list.extend(sm)  # Append smoothed data
            # Check the length match
            if len(sm_list) != len(df):
                raise ValueError(f"Length mismatch for column {col}: sm_list={len(sm_list)}, df={len(df)}")
            # Add smoothed column to DataFrame
            df[col + '_smoothed'] = sm_list
    return df

# Apply smoothing to training and test sets
df_train = smooth_series(df_train, beta)
df_test = smooth_series(df_test, beta)

df_train.columns

Index(['unit', 'time', 'os1', 'os2', 's2', 's3', 's4', 's6', 's7', 's8', 's9',
       's11', 's12', 's13', 's14', 's15', 's17', 's20', 's21', 'rul',
       'os1_smoothed', 'os2_smoothed', 's2_smoothed', 's3_smoothed',
       's4_smoothed', 's6_smoothed', 's7_smoothed', 's8_smoothed',
       's9_smoothed', 's11_smoothed', 's12_smoothed', 's13_smoothed',
       's14_smoothed', 's15_smoothed', 's17_smoothed', 's20_smoothed',
       's21_smoothed'],
      dtype='object')

# Set scientific theme
sns.set_theme(style="whitegrid", font="Times New Roman")

# Select unit
sample_unit = 10
sample_df = df_train[df_train['unit'] == sample_unit].copy()

# Extract sensor columns
sensor_cols = [col for col in df_train.columns if col.startswith('s') and '_smoothed' not in col]
n_sensors = len(sensor_cols)

if n_sensors == 0:
    raise ValueError("No valid sensor columns found in the dataset.")

# Define layout for compact visualization
num_cols = 5
num_rows = -(-n_sensors // num_cols)

fig, axes = plt.subplots(num_rows, num_cols, figsize=(9, 2.5 * num_rows), sharex=True, constrained_layout=True)

# Flatten axes
axes = axes.flatten()

# Define colors
colors = {"original": "tab:blue", "smoothed": "#D55E00", "line": "black"}

for j, col in enumerate(sensor_cols):
    axes[j].plot(sample_df['time'], sample_df[col], color=colors["original"], label="Original", lw=0.8)
    
    smoothed_col = f"{col}_smoothed"
    if smoothed_col in sample_df.columns:
        axes[j].plot(sample_df['time'], sample_df[smoothed_col], color=colors["smoothed"], label="Smoothed", lw=0.8)

    # Grid and Labels
    axes[j].set_title(col, fontsize=10)
    axes[j].tick_params(axis='both', labelsize=8)
    axes[j].grid(True, linestyle="--", linewidth=0.5, alpha=0.6)

# Hide unused subplots
for ax in axes[n_sensors:]:
    ax.set_visible(False)

# Global legend for clarity
fig.legend(["Original", "Smoothed"], loc="upper center", ncol=2, bbox_to_anchor=(0.5, -0.02), fontsize=9, frameon=False)

# Save and display
output_dir = "figures"
os.makedirs(output_dir, exist_ok=True)
plt.savefig(os.path.join(output_dir, f"Smoothing_Sensors_{file_name}.png"), dpi=600, bbox_inches="tight", pad_inches=0.05)
plt.show()

#Remove the original series

for c in df_train.columns:
    if ('s' in c) and ('smoothed' not in c):
        df_train[c] = df_train[c+'_smoothed']
        df_train.drop(c+'_smoothed', axis = 1, inplace = True)
        
for c in df_test.columns:
    if ('s' in c) and ('smoothed' not in c):
        df_test[c] = df_test[c+'_smoothed']
        df_test.drop(c+'_smoothed', axis = 1, inplace = True)
        
for df in [df_train, df_test]:
    display(df.head())

print('training set time cycles:')
display(df_train.groupby('unit')['time'].max().describe())
print('test set time cycles:')
display(df_test.groupby('unit')['time'].max().describe())

training set time cycles:

count    100.00000
mean     247.20000
std       86.48384
min      145.00000
25%      189.75000
50%      220.50000
75%      279.75000
max      525.00000
Name: time, dtype: float64

test set time cycles:

count    100.000000
mean     165.960000
std       86.891773
min       38.000000
25%      105.000000
50%      148.000000
75%      208.000000
max      475.000000
Name: time, dtype: float64

n_features = len([c for c in df_train.columns if 's' in c]) + 1 #plus one for time
window = 15
print(f'number of features: {n_features}, window size: {window}')

number of features: 18, window size: 15

train_indices = list(train_data[(train_data['rul'] >= (window - 1)) & (train_data['time'] > 10)].index)
val_indices = list(val_data[(val_data['rul'] >= (window - 1)) & (val_data['time'] > 10)].index)

rul_max = max(df_train['rul'])
df_train['rul'] = df_train['rul'] / rul_max 
df_test['rul'] = df_test['rul'] / rul_max 
rul_max

125

# max_time = max(df_train['time'])
df_train_mean = df_train['time'].mean()
df_train_std = df_train['time'].std()
df_train['time'] = (df_train['time'] - df_train_mean) / df_train_std
df_test['time'] = (df_test['time'] - df_train_mean) / df_train_std
display(df_train)

class data(Dataset):
    
    def __init__(self, list_indices, df_train):
        
        self.indices = list_indices
        self.df_train = df_train
        
    def __len__(self):
        
        return len(self.indices)
    
    def __getitem__(self, idx):
        
        ind = self.indices[idx]
        X_ = self.df_train.iloc[ind : ind + window, :].drop(['unit','rul'], axis = 1).copy().to_numpy()
        y_ = self.df_train.iloc[ind + window - 1]['rul']
        
        return X_, y_
    
train = data(train_indices, df_train)
val = data(val_indices, df_train)

trainloader = DataLoader(train, batch_size = 256, shuffle = True)
valloader = DataLoader(val, batch_size = 128, shuffle = True)

units = np.arange(1, no_units+1)

class test(Dataset):
    
    def __init__(self, units, df_test):
        
        self.units = units
        self.df_test = df_test
        
    def __len__(self):
        
        return len(self.units)
    
    def __getitem__(self, idx):
        
        n = self.units[idx]
        U = self.df_test[self.df_test['unit'] == n].copy()
        X_ = U.reset_index().iloc[-window:,:].drop(['index', 'unit','rul'], axis = 1).copy().to_numpy()
        y_ = U['rul'].min()
        
        return X_, y_
    
test = test(units, df_test)
testloader = DataLoader(test, batch_size = 100)

dataiter = iter(trainloader)
x,y = next(dataiter)
x.shape

torch.Size([256, 15, 18])

def anderson(f, x0, m=5, lam=1e-4, max_iter=50, tol=1e-2, beta=1.0, verbose=False):
    """Improved Anderson acceleration for fixed-point iteration."""
    # Shape parameters
    batch, channels, dim = x0.shape
    
    # Storage for historical X and F
    X = torch.zeros(batch, m, channels * dim, dtype=x0.dtype, device=x0.device)
    F = torch.zeros(batch, m, channels * dim, dtype=x0.dtype, device=x0.device)

    # Initialize X and F with the first two iterations
    X[:, 0], F[:, 0] = x0.view(batch, -1), f(x0).view(batch, -1)
    X[:, 1], F[:, 1] = F[:, 0], f(F[:, 0].view(batch, channels, dim)).view(batch, -1)

    # Prepare the H matrix and y vector for Anderson acceleration
    H = torch.zeros(batch, m + 1, m + 1, dtype=x0.dtype, device=x0.device)
    H[:, 0, 1:] = H[:, 1:, 0] = 1  # First row and column for normalization
    y = torch.zeros(batch, m + 1, 1, dtype=x0.dtype, device=x0.device)
    y[:, 0, 0] = 1  # y vector for the Anderson step

    res = []  # To store residuals

    for k in range(2, max_iter):
        n = min(k, m)  # Number of iterations to consider

        # Compute residual matrix G
        G = F[:, :n] - X[:, :n]

        # Compute the H matrix (Gramian + regularization)
        GTG = torch.bmm(G, G.transpose(1, 2))
        H[:, 1:n+1, 1:n+1] = GTG + lam * torch.eye(n, dtype=x0.dtype, device=x0.device)[None]

        # Solve for alpha using least squares
        try:
            alpha = torch.linalg.solve(H[:, :n+1, :n+1], y[:, :n+1])[:, 1:n+1, 0]  # (batch x n)
        except RuntimeError as e:
            if verbose:
                print(f"Solver failed at iteration {k}: {e}")
            break

        # Update X and F
        X_update = beta * (alpha[:, None] @ F[:, :n])[:, 0] + (1 - beta) * (alpha[:, None] @ X[:, :n])[:, 0]
        idx = k % m
        X[:, idx] = X_update
        F[:, idx] = f(X[:, idx].view(batch, channels, dim)).view(batch, -1)

        # Compute residuals
        residual = (F[:, idx] - X[:, idx]).norm(dim=1) / (1e-5 + X[:, idx].norm(dim=1).clamp(min=1e-5))
        res.append(residual.mean().item())  # Mean residual across batch

        if verbose:
            print(f"Iteration {k}, residual: {res[-1]:.6f}")

        # Check for convergence
        if res[-1] < tol:
            if verbose:
                print(f"Converged at iteration {k}, residual: {res[-1]:.6f}")
            break

    return X[:, idx].view(batch, channels, dim), (res, k)

import torch.autograd as autograd

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print('Using device:', device)

class DEQFixedPoint(nn.Module):
    def __init__(self, f, solver, **kwargs):
        super().__init__()
        self.f = f
        self.solver = solver
        self.kwargs = kwargs

    def forward(self, x):
        # Compute forward pass and re-engage autograd tape
        b, c, f = x.shape
        with torch.no_grad():  # Ensure no gradients for intermediate calculations
            z, self.stats = self.solver(lambda z: self.f(z, x), torch.zeros((b, 4 * c, f), device=device), **self.kwargs)

        z = self.f(z, x)

        # Set z to require gradients before the hook
        z = z.requires_grad_()  # Ensure z requires gradients

        # Set up Jacobian vector product (without additional forward calls)
        z0 = z.clone().detach().requires_grad_()  # Ensure z0 requires gradients
        f0 = self.f(z0, x)

        def backward_hook(grad):
            # Perform the backward pass using the solver
            g, self.backward_res = self.solver(
                lambda y: autograd.grad(f0, z0, y, retain_graph=True)[0] + grad,
                grad, **self.kwargs)
            return g

        # Register the backward hook
        z.register_hook(backward_hook)
        return z, self.stats

Using device: cuda

import torch.optim.lr_scheduler as lr_scheduler

# class ScaledLeakyReLU(nn.Module):
#     def __init__(self, negative_slope=0.1, scaled_slope=0.9):
#         super(ScaledLeakyReLU, self).__init__()
#         self.negative_slope = negative_slope
#         self.scaled_slope = scaled_slope

#     def forward(self, x):
#         return torch.where(x > 0, self.scaled_slope * x, self.negative_slope * x)

class DualInputAttention(nn.Module):
    def __init__(self, channels, feature_size):
        super(DualInputAttention, self).__init__()
        self.feature_size = feature_size
        self.query_proj = nn.Linear(self.feature_size, self.feature_size)
        self.key_proj = nn.Linear(self.feature_size, self.feature_size)
        self.value_proj = nn.Linear(self.feature_size, self.feature_size)
        self.output_proj = nn.Linear(self.feature_size, self.feature_size)
        self.query_proj.weight.data.normal_(0, 0.01)
        self.key_proj.weight.data.normal_(0, 0.01)
        self.value_proj.weight.data.normal_(0, 0.01)
        self.output_proj.weight.data.normal_(0, 0.01)
        
    def forward(self, input1, input2):
        """
        Args:
            input1: Tensor of shape (batch, channels, feature_size)
            input2: Tensor of shape (batch, channels, feature_size)
        
        Returns:
            Tensor of shape (batch, channels, feature_size), attended features.
        """
        # Project inputs
        query = self.query_proj(input1)  # (batch, channels, feature_size)
        key = self.key_proj(input2)     # (batch, channels, feature_size)
        value = self.value_proj(input2) # (batch, channels, feature_size)
        
        # Compute similarity (scaled dot product)
        scores = torch.einsum('bci,bcj->bcij', query, key)  # (batch, channels, feature_size, feature_size)
        scores = scores / (self.feature_size ** 0.5)
        
        # Compute attention weights
        attention_weights = F.softmax(scores, dim=-1)  # (batch, channels, feature_size, feature_size)
        
        # Compute attended values
        attended = torch.einsum('bcij,bcj->bci', attention_weights, value)  # (batch, channels, feature_size)
        
        # Combine with input1
        output = self.output_proj(attended + input1)  # (batch, channels, feature_size)
        
        return output


class DELayer(nn.Module):
    def __init__(self, n_channels, feature_size, kernel_size=3):
        super().__init__()
        num_groups = 4
        self.conv0 = nn.Conv1d(n_channels, 2 * n_channels, kernel_size=1, padding='same', bias=False)
        self.conv1 = nn.Conv1d(2 * n_channels, 2 * n_channels, kernel_size=kernel_size, padding='same', bias=False)
        self.conv2 = nn.Conv1d(2 * n_channels, 2 * n_channels, kernel_size=kernel_size, padding='same', bias=False)
        self.norm1 = nn.GroupNorm(num_groups, 4 * n_channels)
        self.norm2 = nn.GroupNorm(num_groups, 4 * n_channels)
        # self.norm3 = nn.GroupNorm(num_groups, 2 * n_channels)
        # self.scaled_leaky_relu = ScaledLeakyReLU(negative_slope=0.1, scaled_slope=0.9)
        self.attention = DualInputAttention(4 * n_channels, feature_size)
        self.conv0.weight.data.normal_(0, 0.01)
        self.conv1.weight.data.normal_(0, 0.01)
        self.conv2.weight.data.normal_(0, 0.01)

    def forward(self, z0, x):
        # z = self.conv2(self.norm1(F.relu(self.conv1(z0))))
        # out = self.attention (self.norm2(z),self.norm3(x))
        # z = self.norm4(F.relu(out))
        x = F.relu(self.conv0(x))
        x_conv1 = F.relu(self.conv1(x))
        x_conv2 = F.relu(self.conv2(x_conv1))
        x_concat = torch.cat([x_conv1, x_conv2], dim=1)
        x = self.norm1(x_concat)
        # z = self.norm2(z0)
        z = self.norm2(F.relu(self.attention(z0, x)))
        return z

# Define the Bayesian Neural Network with Dropout 
class BNN(nn.Module):
    def __init__(self, initial_channels, hidden_dim, dropout):
        super(BNN, self).__init__()
        self.dropout_value = dropout
        self.fc1 = nn.Linear(4 * initial_channels, 8* hidden_dim)
        self.dropout1 = nn.Dropout(self.dropout_value)  # Dropout for Monte Carlo Dropout
        self.fc2 = nn.Linear(8 * hidden_dim, 4 * hidden_dim)
        self.dropout2 = nn.Dropout(self.dropout_value)
        self.fc3 = nn.Linear(4 * hidden_dim, 1)

    def forward(self, x, training=True):
        x = torch.relu(self.fc1(x))
        x = self.dropout1(x) if training else x   # Apply dropout during training and inference
        x = torch.relu(self.fc2(x))
        x = self.dropout2(x) if training else x   # Apply dropout during training and inference
        return self.fc3(x)

class RegressorModel(nn.Module):
    def __init__(self, initial_channels, feature_size, hidden_dim, dropout):
        super(RegressorModel, self).__init__()

        # Layers
        # self.conv =nn.Conv1d(initial_channels, 2 * initial_channels, kernel_size=1, bias=False)
        # self.norm = nn.GroupNorm(4, 2 * initial_channels)
        self.f = DELayer(n_channels= initial_channels, feature_size=feature_size )
        self.DEQ = DEQFixedPoint(self.f, anderson, tol=1e-4, max_iter=200, beta=1.0)
        self.pool = nn.AdaptiveAvgPool1d(1)
        self.feedforward = BNN(initial_channels, hidden_dim, dropout)

    def forward(self, x, training=True):
        # x = self.norm(F.relu(self.conv(x)))
        x = self.DEQ(x)
        x = self.pool(x[0]).squeeze(-1)             # Apply Deep Equilibrium Model
        # x = torch.flatten(x, start_dim=1)
        output = self.feedforward(x, training)            # Apply feedforward Model for regression
  
        return output

    
learning_rate = 1e-3 
init_channels = window
inner_channels = window
n_hidden_units =  window
set_dropout = 0.4

model = RegressorModel(initial_channels=init_channels, feature_size=n_features, hidden_dim=n_hidden_units, dropout=set_dropout ).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) 
scheduler = lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.5)

X= torch.randn((64, window, n_features), device=device)
f = DELayer(n_channels=init_channels, feature_size= n_features).to(device)
DEQ = DEQFixedPoint(f, anderson, tol=1e-4, max_iter=100, beta=0.9).to(device)
out = DEQ(X)[0]
print(out.shape)
out2 = nn.AdaptiveAvgPool1d(1)(out).squeeze(-1)
print(out2.shape)
out3 = model(X)[0]
print(out3.shape)

torch.Size([64, 60, 18])
torch.Size([64, 60])
torch.Size([1])

# def validation(loss_fn, num_samples=100):
#     model.train()  # Enable stochastic behavior (dropout during inference)
#     total_loss = 0.0
#     total_samples = 0

#     with torch.no_grad():
#         for X, y in valloader:
#             X, y = X.to(device).float(), y.to(device).float()

#             stochastic_predictions = []

#             # Perform stochastic forward passes
#             for _ in range(num_samples):
#                 y_pred = model(X, training=True).squeeze()
#                 stochastic_predictions.append(y_pred.cpu().numpy())

#             # Convert to NumPy array
#             stochastic_predictions = np.array(stochastic_predictions)  # Shape: (num_samples, batch_size)

#             # Calculate mean prediction
#             mean_pred = torch.tensor(np.mean(stochastic_predictions, axis=0)).to(device)

#             # Compute batch loss
#             loss = loss_fn(mean_pred, y)
#             total_loss += loss.item() * X.size(0)  # Accumulate the weighted batch loss
#             total_samples += X.size(0)

#     avg_val_loss = total_loss / total_samples  # Calculate average loss
#     return avg_val_loss

def validation(loss_fn):
    model.eval()  # Set the model to evaluation mode
    total_loss = 0.0
    total_samples = 0

    with torch.no_grad():  # Disable gradient computation for validation
        for X, y in valloader:  # Iterate over all batches in the validation loader
            X, y = X.to(device).to(torch.float32), y.to(device).to(torch.float32)
            
            # Use enable_grad for specific models if required
            if isinstance(model, RegressorModel):  # Replace with the DEQ model class
                with torch.enable_grad():
                    y_pred = model(X).squeeze()
            else:
                y_pred = model(X).squeeze()
            
            y = y.squeeze()
            loss = loss_fn(y_pred, y)
            total_loss += loss.item() * X.size(0)  # Accumulate the weighted batch loss
            total_samples += X.size(0)  # Count the total samples processed

    avg_val_loss = total_loss / total_samples  # Calculate average loss
    return avg_val_loss

loss_L1 = nn.L1Loss()
    
def test():
    model.train()  # Enable stochastic behavior (dropout during inference)
    
    total_loss_MSE = 0.0
    total_loss_L1 = 0.0
    total_ASUE = 0.0
    all_stochastic_predictions = []
    all_pred_values = []
    all_true_values = []
    all_uncertainties = []  # Collect uncertainties
    num_samples = 100

    with torch.no_grad():
        for X, y in testloader:
            X, y = X.to(device).float(), y.to(device).float()

            stochastic_predictions = []

            # Perform stochastic forward passes
            for _ in range(num_samples):
                y_pred = model(X, training=True).squeeze()
                stochastic_predictions.append(y_pred.cpu().numpy())
            
            # Convert to NumPy array
            stochastic_predictions = np.array(stochastic_predictions)  # Shape: (num_samples, batch_size)

            # Calculate mean and uncertainty
            mean_pred = torch.tensor(np.mean(stochastic_predictions, axis=0)).to(device)
            uncertainty_batch = np.var(stochastic_predictions, axis=0)  # Variance as uncertainty

            # Compute batch losses
            loss_MSE = torch.mean((mean_pred - y) ** 2).item()
            loss_L1_val = loss_L1(mean_pred, y).item()
            ASUE = torch.mean(torch.relu(y - mean_pred)).item()

            # Aggregate results
            total_loss_MSE += loss_MSE * X.size(0)
            total_loss_L1 += loss_L1_val * X.size(0)
            total_ASUE += ASUE * X.size(0)

            # Collect predictions, true values, and uncertainties
            all_stochastic_predictions.append(stochastic_predictions)
            all_pred_values.append(mean_pred.cpu().numpy())
            all_true_values.append(y.cpu().numpy())
            all_uncertainties.append(uncertainty_batch)
        
    # Average metrics
    total_samples = len(testloader.dataset)
    avg_loss_MSE = total_loss_MSE / total_samples
    avg_loss_L1 = total_loss_L1 / total_samples
    avg_ASUE = total_ASUE / total_samples

    # Concatenate all predictions, uncertainties, and true values
    all_pred_values = np.concatenate([x.flatten() for x in all_pred_values], axis=0)
    all_stochastic_predictions = np.concatenate(all_stochastic_predictions, axis=1)
    all_true_values = np.concatenate(all_true_values, axis=0)
    all_uncertainties = np.concatenate(all_uncertainties, axis=0)  # Ensure consistency

    return avg_loss_MSE, avg_loss_L1, avg_ASUE, all_uncertainties, all_stochastic_predictions, all_pred_values, all_true_values

# Initialize lists for tracking losses
T, V = [], []
epochs = 35

loss_fn = nn.MSELoss()
siterations, sresiduals = [], []
stats = {}

# Define forward hook for DEQ statistics
def forward_hook(module, input, output):
    stats[module] = output

# Register forward hook
hook_handle = model.DEQ.register_forward_hook(forward_hook)

best_val_loss = float('inf')  # Track best validation loss

pbar = tqdm(range(epochs), desc="Training Progress", dynamic_ncols=True)
for epoch in pbar:
    model.train()
    epoch_loss = 0
    sresiduals_list, siterations_list = [], []

    for X, y in trainloader:
        X, y = X.to(device, dtype=torch.float32), y.to(device, dtype=torch.float32)

        # Forward pass
        y_pred = model(X).squeeze()
        y = y.squeeze()
        loss = loss_fn(y_pred, y)
        epoch_loss += loss.item()

        # Backward pass
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        # Collect DEQ statistics  
        iterations = stats.get(model.DEQ, (None, [0, 0]))[1] 
        if iterations and len(iterations) >= 2:
            sresiduals_list.append(iterations[0][-1])
            siterations_list.append(iterations[1])

    # Step the scheduler
    scheduler.step()

    # Validation loss
    val_loss = validation(loss_fn)
    T.append(epoch_loss / len(trainloader))
    V.append(val_loss)
 
    # Update tqdm bar
    pbar.set_postfix({'Train Loss': T[-1], 'Val Loss': val_loss})

    # Track DEQ statistics
    sresiduals.append(np.mean(sresiduals_list) if sresiduals_list else 0)
    siterations.append(np.mean(siterations_list) if siterations_list else 0)

    # Save best model
    if val_loss < best_val_loss:
        best_val_loss = val_loss
        torch.save(model.state_dict(), 'best_model.pth')

    print(f'Epoch {epoch + 1}/{epochs} - Train Loss: {T[-1]:.4f}, Val Loss: {V[-1]:.4f}')

# Clean up hook handle
hook_handle.remove()

print(sresiduals)
print(siterations)

torch.save(model.state_dict(), f'saved_models\\DEM_{file_name}')

# Perform testing with Monte Carlo Dropout
mse, l1, asue, uncertainty, all_y_pred, y_pred, y = test()

# Calculate RMSE and scale by RUL max value
rmse = rul_max * np.sqrt(mse)

# Print results
print(f"Test Results:\n"
      f"RMSE: {round(rmse, 2)}\n"
      f"L1 Loss: {round(l1, 2)}\n"
      f"ASUE: {round(asue, 2)}")

# Mean uncertainty
mean_uncertainty = np.mean(uncertainty)
print(f"Mean Uncertainty: {round(mean_uncertainty, 4)}")

# Optional: Visualize predictions, true values, and uncertainties
import matplotlib.pyplot as plt

plt.figure(figsize=(14, 8))

# True vs predicted RUL
plt.plot(range(len(y)), y, label="True Values", c="lightseagreen", marker=".", alpha=0.7)
plt.plot(range(len(y_pred)), y_pred, label="Predicted Values", c="salmon", marker=".", alpha=0.7)

# Uncertainty as error bars
plt.fill_between(
    range(len(y_pred)),
    y_pred - np.sqrt(uncertainty),
    y_pred + np.sqrt(uncertainty),
    color="salmon", alpha=0.2, label="Uncertainty"
)

# Add labels and legend
plt.xlabel("Sample Index", fontsize=14)
plt.ylabel("RUL", fontsize=14)
plt.title("True vs Predicted RUL with Uncertainty", fontsize=16)
plt.legend(fontsize=12)
plt.grid(True)
plt.savefig(f'figures\\Responses_with_Uncertainty_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()

print(y)
print(y_pred)
print(1*(y_pred>y))
print(all_y_pred.shape)

print(all_y_pred)

# Create a figure and axes
fig, ax = plt.subplots(1, 1, figsize=(14, 8))

# Plot predictions
ax.plot(np.arange(1, y_pred.shape[0] + 1), y_pred, label='Predictions', color='salmon', marker='.')

# Plot true values
ax.plot(np.arange(1, y.shape[0] + 1), y, label='True Values', color='lightseagreen', marker='.')

# Set limits, labels, grid, and legend
ax.set_ylim([0, 1])
ax.set_xlabel('Test Engine Units', fontsize=16)
ax.set_ylabel('RUL', fontsize=16)
ax.grid(True)
ax.legend()
plt.savefig(f'figures\\Responses_without_Uncertainty_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
# Show the plot
plt.show()

plt.plot(np.arange(1,len(T)+1), T, label= 'Train loss')
plt.plot(np.arange(1,len(V)+1), V, label = 'Validation loss')
plt.legend()
plt.grid()
plt.xlabel('Epochs')
plt.ylabel('MSE')
plt.savefig(f'figures\\Training_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show

fig, ax = plt.subplots(figsize = (12,8))

def animate(i):
    ax.clear()
    line1, = ax.plot(np.arange(1,i+1), T[:i], label = 'train_loss')
    line2, = ax.plot(np.arange(1,i+1), V[:i], label = 'val_loss')
    ax.legend()
    ax.grid(True)
    ax.set_xlim(0,101)
    ax.set_ylim(0,4000)
    ax.set_xlabel('epochs')
    ax.set_ylabel('MSE')
                     
    return line1, line2
                     
    
animation = FA(fig, animate, np.arange(1,len(T)+1), interval = 50)

%time animation.save('animation3.gif', writer='imagemagick', fps=20)

plt.close(fig)

# Count total parameters
total_params = sum(p.numel() for p in model.parameters())
print(f"Total number of parameters: {total_params}")

# Count trainable parameters only
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"Trainable parameters: {trainable_params}")

from torchinfo import summary
summary(model, input_size=(64, window, n_features))

model.load_state_dict(torch.load(f'saved_models\\DEM_{file_name}', map_location=torch.device(device)))

def calculate_phm08_score(true_rul, predicted_rul, alpha=13, beta=10):
    """
    Calculate the PHM08 score metric for RUL prediction.

    Parameters:
        true_rul (array-like): Array of true RUL values.
        predicted_rul (array-like): Array of predicted RUL values.
        alpha (float): Scaling factor for early predictions (default=13).
        beta (float): Scaling factor for late predictions (default=10).

    Returns:
        float: Total PHM08 score.
    """
    true_rul = np.array(true_rul)
    predicted_rul = np.array(predicted_rul)

    if true_rul.shape != predicted_rul.shape:
        raise ValueError("Shape mismatch between true and predicted RUL arrays")

    d = (predicted_rul - true_rul) * MAX_RUL

    score = np.where(d < 0, np.exp(-d / alpha) - 1, np.exp(d / beta) - 1)
    return np.sum(score)

num_samples = 100

def test(num_samples=num_samples):
    model.train()  # Enable stochastic behavior (dropout during inference)
    
    total_loss_MSE = 0.0
    total_loss_L1 = 0.0
    total_ASUE = 0.0
    total_phm08_score = 0.0
    all_stochastic_predictions = []
    all_pred_values = []
    all_true_values = []
    all_uncertainties = []  # Collect uncertainties

    with torch.no_grad():
        for X, y in testloader:
            X, y = X.to(device).float(), y.to(device).float()

            stochastic_predictions = []

            # Perform stochastic forward passes
            for _ in range(num_samples):
                y_pred = model(X, training=True).squeeze()
                stochastic_predictions.append(y_pred.cpu().numpy())
            
            # Convert to NumPy array
            stochastic_predictions = np.array(stochastic_predictions)  # Shape: (num_samples, batch_size)

            # Calculate mean and uncertainty with std
            mean_pred = torch.tensor(np.mean(stochastic_predictions, axis=0)).to(device)
            uncertainty_batch = np.std(stochastic_predictions, axis=0)  # Variance as uncertainty

            # Compute batch losses
            loss_MSE = torch.mean((mean_pred - y) ** 2).item()
            loss_L1_val = loss_L1(mean_pred, y).item()
            ASUE = torch.mean(torch.relu(y - mean_pred)).item()

            # Compute PHM08 score
            phm08_score = calculate_phm08_score(y.cpu().numpy(), mean_pred.cpu().numpy())

            # Aggregate results
            total_loss_MSE += loss_MSE * X.size(0)
            total_loss_L1 += loss_L1_val * X.size(0)
            total_ASUE += ASUE * X.size(0)
            total_phm08_score += phm08_score

            # Collect predictions, true values, and uncertainties
            all_stochastic_predictions.append(stochastic_predictions)
            all_pred_values.append(mean_pred.cpu().numpy())
            all_true_values.append(y.cpu().numpy())
            all_uncertainties.append(uncertainty_batch)
        
    # Average metrics
    total_samples = len(testloader.dataset)
    avg_loss_MSE = total_loss_MSE / total_samples
    avg_loss_L1 = total_loss_L1 / total_samples
    avg_ASUE = total_ASUE / total_samples
    avg_phm08_score = total_phm08_score

    # Concatenate all predictions, uncertainties, and true values
    all_pred_values = np.concatenate([x.flatten() for x in all_pred_values], axis=0)
    all_stochastic_predictions = np.concatenate(all_stochastic_predictions, axis=1)
    all_true_values = np.concatenate(all_true_values, axis=0)
    all_uncertainties = np.concatenate(all_uncertainties, axis=0)  # Ensure consistency

    return avg_loss_MSE, avg_loss_L1, avg_ASUE, avg_phm08_score, all_uncertainties, all_stochastic_predictions, all_pred_values, all_true_values

# Perform testing with Monte Carlo Dropout
mse, l1, asue, score, uncertainty, all_y_pred, y_pred, y = test()

# Calculate RMSE and scale by RUL max value
rmse = rul_max * np.sqrt(mse)

# Print results
print(f"Test Results:\n"
      f"RMSE: {round(rmse, 2)}\n"
      f"Score: {round(score, 2)}\n"
      f"L1 Loss: {round(l1, 2)}\n"
      f"ASUE: {round(asue, 2)}")

# Mean uncertainty
mean_uncertainty = np.mean(uncertainty)
print(f"Mean Uncertainty: {round(mean_uncertainty, 4)}")

# Optional: Visualize predictions, true values, and uncertainties
import matplotlib.pyplot as plt

plt.figure(figsize=(14, 8))

# True vs predicted RUL
plt.plot(range(len(y)), y, label="True Values", c="lightseagreen", marker=".", alpha=0.7)
plt.plot(range(len(y_pred)), y_pred, label="Predicted Values", c="salmon", marker=".", alpha=0.7)

# Uncertainty as error bars
plt.fill_between(
    range(len(y_pred)),
    y_pred - uncertainty,
    y_pred + uncertainty,
    color="salmon", alpha=0.2, label="Uncertainty"
)

# Add labels and legend
plt.xlabel("Sample Index", fontsize=14)
plt.ylabel("RUL", fontsize=14)
plt.title("True vs Predicted RUL with Uncertainty", fontsize=16)
plt.legend(fontsize=12)
plt.grid(True)
plt.savefig(f'figures\\Responses_with_Uncertainty_for_dataset_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()

values = np.array([y, y_pred, uncertainty]).T
df_values = pd.DataFrame(values)
values = df_values.sort_values(0, ascending=False).to_numpy()

y, y_pred, uncertainty = np.array(values).T

# y, y_pred, uncertainty = MAX_RUL * np.array(values).T
# uncertainty = MAX_RUL * uncertainty

plt.hist(uncertainty, bins=50, color="blue", alpha=0.7, label="Uncertainty Distribution")
plt.xlabel("Uncertainty")
plt.ylabel("Frequency")
plt.title("Histogram of Prediction Uncertainty")
plt.legend()
plt.grid(True)
plt.show()

# confidence = 0.99  # Confidence level
# zeta = 2.576

# # Calculate confidence interval
# standard_error = uncertainty / np.sqrt(num_samples)  # Standard error of the mean
# margin_of_error = zeta * standard_error  # Margin of error

confidence = 0.95  # Confidence level
margin_of_error = confidence * uncertainty  # Margin of error

plt.figure(figsize=(8, 6)) # Adjust figure size for publication

# True Values (more prominent line)
plt.plot(range(len(y)), y, label="True RUL", c="lightseagreen", linewidth=2)

# Predicted Values (smaller scatter points)
plt.scatter(range(len(y_pred)), y_pred, label="Predicted RUL", c="salmon", edgecolor="black", alpha=0.7, zorder=3, s=15) # reduced scatter point size to 15

# Vertical Lines (improved clarity)
for i in range(len(y)):
    plt.plot([i, i], [y_pred[i], y[i]], c="gray", linestyle="--", linewidth=0.8, alpha=0.7)

# Confidence Interval (adjusted opacity, more descriptive label)
plt.fill_between(
    range(len(y_pred)),
    y_pred - margin_of_error,
    y_pred + margin_of_error,
    color="salmon",
    alpha=0.3,  # Reduced opacity
    label=f"{int(confidence * 100)}% Prediction Interval"  # More accurate label
)

# Labels and Title (improved clarity)
dataset_name = file_name[:-4]
plt.xlabel("Testing Sample Index") # Removed unnecessary context from label
plt.ylabel("Remaining Useful Life (RUL)") # Removed unnecessary context from label
plt.title(f"Comparison of True and Predicted RUL for {dataset_name}") # Removed unnecessary context from title

# Legend and Grid
plt.legend()

# Save and Show
plt.savefig(f'figures/True_vs_PredictedRUL_{file_name}.png')
plt.show()

# Calculate errors and mean predictions
errors = np.abs(y_pred - y)

# Scatter plot with enhancements
plt.figure(figsize=(8, 6))
sc = plt.scatter(uncertainty, errors, c=errors, cmap='viridis', alpha=0.7, edgecolor='k')

# Add colorbar
cbar = plt.colorbar(sc)
cbar.set_label('Error Magnitude')

# Add grid and labels
plt.grid(alpha=0.3)
plt.xlabel('Uncertainty', fontsize=12)
plt.ylabel('Error', fontsize=12)
plt.title('Uncertainty vs. Error', fontsize=14)

# Polynomial fit (degree 2 or higher)
degree = 1  # try 3, 4, etc. for more complex trends
z = np.polyfit(uncertainty, errors, degree)
p = np.poly1d(z)

# Sort for smooth plotting
sorted_idx = np.argsort(uncertainty)
plt.plot(uncertainty[sorted_idx], p(uncertainty[sorted_idx]),
         color='red', linestyle='--', label=f'Polynomial Regression (deg {degree})')

# plt.plot(uncertainty, p(uncertainty), color='red', linestyle='--', label='Trend Line')

# Add legend
plt.legend()

# Show the plot
plt.tight_layout()
plt.savefig(f'figures\\UncertaintyVSError_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()


# Compute statistics
correlation = np.corrcoef(uncertainty, errors)[0, 1]
print(f"Correlation between uncertainty and error: {correlation:.2f}")

# Evaluate performance in high/low uncertainty regions
threshold = np.percentile(uncertainty, 75)  # Upper 25% uncertainty
high_uncertainty_mask = uncertainty > threshold
low_uncertainty_mask = uncertainty <= threshold

high_uncertainty_error = errors[high_uncertainty_mask].mean()
low_uncertainty_error = errors[low_uncertainty_mask].mean()

print(f"Mean error in high uncertainty regions: {high_uncertainty_error:.2f}")
print(f"Mean error in low uncertainty regions: {low_uncertainty_error:.2f}")

# Calculate absolute error
errors = np.abs(y_pred - y)

# Plot
plt.figure(figsize=(8, 6))
sc = plt.scatter(y, errors, c=errors, cmap='magma', alpha=0.7, edgecolor='k')

# Add colorbar
cbar = plt.colorbar(sc)
cbar.set_label('Error Magnitude')

# Add labels and title
plt.xlabel('True Values', fontsize=12)
plt.ylabel('Absolute Error', fontsize=12)
plt.title('True Values vs. Prediction Error', fontsize=14)
plt.grid(alpha=0.3)

# Optional trend line
z = np.polyfit(y, errors, 2)
p = np.poly1d(z)
sorted_idx = np.argsort(y)
plt.plot(y[sorted_idx], p(y[sorted_idx]), color='red', linestyle='--', label='Trend (deg 2)')

# Add legend
plt.legend()

# Save and show
plt.tight_layout()
plt.savefig(f'figures\\TrueValueVsError_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()

model.load_state_dict(torch.load(f'saved_models\\DEM_{file_name}', map_location=torch.device(device)))

def calculate_phm08_score(true_rul, predicted_rul, alpha=13, beta=10):
    """
    Calculate the PHM08 score metric for RUL prediction.

    Parameters:
        true_rul (array-like): Array of true RUL values.
        predicted_rul (array-like): Array of predicted RUL values.
        alpha (float): Scaling factor for early predictions (default=13).
        beta (float): Scaling factor for late predictions (default=10).

    Returns:
        float: Total PHM08 score.
    """
    true_rul = np.array(true_rul)
    predicted_rul = np.array(predicted_rul)

    if true_rul.shape != predicted_rul.shape:
        raise ValueError("Shape mismatch between true and predicted RUL arrays")

    d = (predicted_rul - true_rul) * MAX_RUL

    score = np.where(d < 0, np.exp(-d / alpha) - 1, np.exp(d / beta) - 1)
    return np.sum(score)

def test_without_dropout():
    model.eval()  # Disable stochastic behavior (dropout during inference)

    total_loss_MSE = 0.0
    total_loss_L1 = 0.0
    total_ASUE = 0.0
    total_phm08_score = 0.0
    all_pred_values = []
    all_true_values = []

    with torch.no_grad():
        for X, y in testloader:
            X, y = X.to(device).float(), y.to(device).float()

            # Perform a single deterministic forward pass
            y_pred = model(X).squeeze()

            # Compute batch losses
            loss_MSE = torch.mean((y_pred - y) ** 2).item()
            loss_L1_val = loss_L1(y_pred, y).item()
            ASUE = torch.mean(torch.relu(y - y_pred)).item()

            # Compute PHM08 score
            phm08_score = calculate_phm08_score(y.cpu().numpy(), y_pred.cpu().numpy())

            # Aggregate results
            total_loss_MSE += loss_MSE * X.size(0)
            total_loss_L1 += loss_L1_val * X.size(0)
            total_ASUE += ASUE * X.size(0)
            total_phm08_score += phm08_score

            # Collect predictions and true values
            all_pred_values.append(y_pred.cpu().numpy())
            all_true_values.append(y.cpu().numpy())

    # Average metrics
    total_samples = len(testloader.dataset)
    avg_loss_MSE = total_loss_MSE / total_samples
    avg_loss_L1 = total_loss_L1 / total_samples
    avg_ASUE = total_ASUE / total_samples
    avg_phm08_score = total_phm08_score

    # Concatenate all predictions and true values
    all_pred_values = np.concatenate([x.flatten() for x in all_pred_values], axis=0)
    all_true_values = np.concatenate(all_true_values, axis=0)

    return avg_loss_MSE, avg_loss_L1, avg_ASUE, avg_phm08_score, all_pred_values, all_true_values

# Perform testing without Monte Carlo Dropout
mse, l1, asue, score, y_pred, y = test_without_dropout()

# Calculate RMSE and scale by RUL max value
rmse = rul_max * np.sqrt(mse)

# Print results
print(f"Test Results:\n"
      f"RMSE: {rmse:.2f}\n"
      f"Score: {score:.2f}\n"
      f"L1 Loss: {l1:.2f}\n"
      f"ASUE: {asue:.3f}")

# Perform testing with Monte Carlo Dropout
mse, l1, asue, score, uncertainty, all_y_pred, y_pred, y = test()

loss_MSE = np.mean((y_pred - y) ** 2).item()
rmse = rul_max * np.sqrt(mse)
print(f'RMSE:{rmse:.2f}')

all_y_pred_mean = all_y_pred.mean(axis=0)
weights = abs(all_y_pred - all_y_pred_mean)
weights = np.exp(-weights)
# weights = 1 / weights
weights /= weights.sum(axis=0)
weighted_predictions = np.sum(weights * all_y_pred, axis=0)

mse = np.mean((weighted_predictions - y) ** 2).item()
rmse = rul_max * np.sqrt(mse)
print(f'RMSE:{rmse:.2f}')

# Define a threshold based on deviations (e.g., 1.5x the mean deviation)
threshold = 1.5 * weights.mean(axis=0)

# Mask predictions with deviations exceeding the threshold
valid_predictions = np.where(weights <= threshold, all_y_pred, np.nan)

# Compute predictions as the mean of valid predictions
weighted_predictions = np.nanmean(valid_predictions, axis=0)

mse = np.mean((weighted_predictions - y) ** 2).item()
rmse = rul_max * np.sqrt(mse)
print(f'RMSE:{rmse:.2f}')

model.load_state_dict(torch.load(f'saved_models\\DEM_{file_name}', map_location=torch.device(device)))

C:\Users\spiro\AppData\Local\Temp\ipykernel_22120\4004060783.py:1: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  model.load_state_dict(torch.load(f'saved_models\\DEM_{file_name}', map_location=torch.device(device)))

<All keys matched successfully>

def calculate_phm08_score(true_rul, predicted_rul, alpha=13, beta=10):
    """
    Calculate the PHM08 score metric for RUL prediction.

    Parameters:
        true_rul (array-like): Array of true RUL values.
        predicted_rul (array-like): Array of predicted RUL values.
        alpha (float): Scaling factor for early predictions (default=13).
        beta (float): Scaling factor for late predictions (default=10).

    Returns:
        float: Total PHM08 score.
    """
    true_rul = np.array(true_rul)
    predicted_rul = np.array(predicted_rul)

    if true_rul.shape != predicted_rul.shape:
        raise ValueError("Shape mismatch between true and predicted RUL arrays")

    d = (predicted_rul - true_rul) * MAX_RUL

    score = np.where(d < 0, np.exp(-d / alpha) - 1, np.exp(d / beta) - 1)
    return np.sum(score)

num_samples = 100

def test(num_samples=num_samples):
    model.train()  # Enable stochastic behavior (dropout during inference)
    
    total_loss_MSE = 0.0
    total_loss_L1 = 0.0
    total_ASUE = 0.0
    total_phm08_score = 0.0
    all_stochastic_predictions = []
    all_pred_values = []
    all_true_values = []
    all_uncertainties = []  # Collect uncertainties

    with torch.no_grad():
        for X, y in testloader:
            X, y = X.to(device).float(), y.to(device).float()

            stochastic_predictions = []

            # Perform stochastic forward passes
            for _ in range(num_samples):
                y_pred = model(X, training=True).squeeze()
                stochastic_predictions.append(y_pred.cpu().numpy())
            
            # Convert to NumPy array
            stochastic_predictions = np.array(stochastic_predictions)  # Shape: (num_samples, batch_size)

            # Calculate mean and uncertainty with std
            mean_pred = torch.tensor(np.mean(stochastic_predictions, axis=0)).to(device)
            uncertainty_batch = np.std(stochastic_predictions, axis=0)  # Variance as uncertainty

            # Compute batch losses
            loss_MSE = torch.mean((mean_pred - y) ** 2).item()
            loss_L1_val = loss_L1(mean_pred, y).item()
            ASUE = torch.mean(torch.relu(y - mean_pred)).item()

            # Compute PHM08 score
            phm08_score = calculate_phm08_score(y.cpu().numpy(), mean_pred.cpu().numpy())

            # Aggregate results
            total_loss_MSE += loss_MSE * X.size(0)
            total_loss_L1 += loss_L1_val * X.size(0)
            total_ASUE += ASUE * X.size(0)
            total_phm08_score += phm08_score

            # Collect predictions, true values, and uncertainties
            all_stochastic_predictions.append(stochastic_predictions)
            all_pred_values.append(mean_pred.cpu().numpy())
            all_true_values.append(y.cpu().numpy())
            all_uncertainties.append(uncertainty_batch)
        
    # Average metrics
    total_samples = len(testloader.dataset)
    avg_loss_MSE = total_loss_MSE / total_samples
    avg_loss_L1 = total_loss_L1 / total_samples
    avg_ASUE = total_ASUE / total_samples
    avg_phm08_score = total_phm08_score

    # Concatenate all predictions, uncertainties, and true values
    all_pred_values = np.concatenate([x.flatten() for x in all_pred_values], axis=0)
    all_stochastic_predictions = np.concatenate(all_stochastic_predictions, axis=1)
    all_true_values = np.concatenate(all_true_values, axis=0)
    all_uncertainties = np.concatenate(all_uncertainties, axis=0)  # Ensure consistency

    return avg_loss_MSE, avg_loss_L1, avg_ASUE, avg_phm08_score, all_uncertainties, all_stochastic_predictions, all_pred_values, all_true_values

def validate_4_calibrate(num_samples=100):
    model.train()  # Dropout needs to stay on
    
    all_pred_values = []
    all_uncertainties = []
    all_true_values = []

    with torch.no_grad():
        for X, y in valloader:
            X, y = X.to(device).float(), y.to(device).float()

            stochastic_predictions = []

            for _ in range(num_samples):
                y_pred = model(X, training=True).squeeze()
                stochastic_predictions.append(y_pred.cpu().numpy())

            stochastic_predictions = np.array(stochastic_predictions)  # (num_samples, batch_size)
            mean_pred = np.mean(stochastic_predictions, axis=0)
            uncertainty = np.std(stochastic_predictions, axis=0)

            all_pred_values.append(mean_pred)
            all_uncertainties.append(uncertainty)
            all_true_values.append(y.cpu().numpy())

    # Flatten
    val_pred_means = np.concatenate(all_pred_values)
    val_pred_stds = np.concatenate(all_uncertainties)
    val_targets = np.concatenate(all_true_values)

    return val_pred_means, val_pred_stds, val_targets



val_pred_means, val_pred_stds, val_targets = validate_4_calibrate(num_samples=100)

mse, l1, asue, score, uncertainty, all_y_pred, y_pred, y = test()

def compute_coverage(y_true, y_pred, scaled_stds):
    """
    Computes the empirical coverage: the fraction of samples where
    the true value is inside the predicted confidence interval.
    
    Parameters:
    - y_true : np.array, ground truth values
    - y_pred : np.array, predicted means
    - scaled_stds : np.array, calibrated standard deviations (after applying k-scaling)

    Returns:
    - coverage : float, percentage of samples inside the confidence interval
    """
    lower_bounds = y_pred - scaled_stds
    upper_bounds = y_pred + scaled_stds

    # Check if the true value lies inside the interval
    inside = (y_true >= lower_bounds) & (y_true <= upper_bounds)

    # Calculate coverage as a percentage
    coverage = np.mean(inside) * 100.0
    return coverage


coverage = compute_coverage(y, y_pred, uncertainty)
print(f"Ratio of coverage:{coverage}%")

Ratio of coverage:63.0%

values = np.array([y, y_pred, uncertainty]).T
df_values = pd.DataFrame(values)
values = df_values.sort_values(0, ascending=False).to_numpy()

y, y_pred, uncertainty = np.array(values).T

# y, y_pred, uncertainty = MAX_RUL * np.array(values).T
# uncertainty = MAX_RUL * uncertainty

import scipy.stats as st

import scipy.optimize as optim

def calibrate_scaling(val_pred_means, val_stds, val_targets, alpha=0.05):
    def coverage_diff(k):
        lower = val_pred_means - k * val_stds
        upper = val_pred_means + k * val_stds
        coverage = np.mean((val_targets >= lower) & (val_targets <= upper))
        return abs((1 - alpha) - coverage)

    result = optim.minimize_scalar(coverage_diff, bounds=(0.1, 5.0), method='bounded')  # Adjust bounds
    return result.x

k_opt = calibrate_scaling(val_pred_means, val_pred_stds, val_targets, alpha=0.05)
display(k_opt)

2.6742587996512373

calibrated_lower = y_pred - k_opt * uncertainty
calibrated_upper = y_pred + k_opt * uncertainty

test_coverage = np.mean((y >= calibrated_lower) & (y <= calibrated_upper))
display(f'Coverage {test_coverage*100}%')

'Coverage 94.0%'

def plot_reliability(pred_means, pred_stds, targets):
    confidences = np.linspace(0.5, 0.99, 20)
    coverages = []

    for conf in confidences:
        z = st.norm.ppf(1 - (1 - conf)/2)
        lower = pred_means - z * pred_stds
        upper = pred_means + z * pred_stds
        coverage = np.mean((targets >= lower) & (targets <= upper))
        coverages.append(coverage)

    plt.plot(confidences, coverages, label='Empirical Coverage')
    plt.plot([0.5, 0.99], [0.5, 0.99], 'k--', label='Ideal Calibration')
    plt.xlabel("Nominal Confidence")
    plt.ylabel("Empirical Coverage")
    plt.legend()
    plt.grid()
    plt.title("Calibration Curve")
    plt.show()

plot_reliability(y_pred, uncertainty, y)

# confidence = 0.99  # Confidence level
# zeta = 2.576

# # Calculate confidence interval
# standard_error = uncertainty / np.sqrt(num_samples)  # Standard error of the mean
# margin_of_error = zeta * standard_error  # Margin of error

confidence = 0.95  # Confidence level
margin_of_error = k_opt * uncertainty  # Margin of error
margin_of_error = confidence * margin_of_error   # Margin of error

# Boolean mask: True if point is inside the confidence interval
inside_interval = (y >= y_pred - margin_of_error) & (y <= y_pred + margin_of_error)

# Assign colors: green if inside, red if outside
colors_interval = np.where(inside_interval, 'mediumseagreen', 'crimson')

plt.figure(figsize=(8, 6)) # Adjust figure size for publication
plt.ylim(0, MAX_RUL+1)

# True Values (more prominent line)
plt.plot(range(len(y)), y * MAX_RUL, label="True RUL", c="lightseagreen", linewidth=2)

# Predicted Values (smaller scatter points)
plt.scatter(range(len(y_pred)), y_pred * MAX_RUL, label="Predicted RUL", c=colors_interval, edgecolor="black", alpha=0.7, zorder=3, s=15) # reduced scatter point size to 15

# Vertical Lines (improved clarity)
for i in range(len(y)):
    plt.plot([i, i], [y_pred[i], y[i]], c="gray", linestyle="--", linewidth=0.8, alpha=0.7)

# Confidence Interval (adjusted opacity, more descriptive label)
plt.fill_between(
    range(len(y_pred)),
    (y_pred - margin_of_error)* MAX_RUL,
    np.minimum(y_pred + margin_of_error, 1) * MAX_RUL,
    color="salmon",
    alpha=0.3,  # Reduced opacity
    label=f"{int(confidence * 100)}% Prediction Interval"  # More accurate label
)

# Labels and Title (improved clarity)
dataset_name = file_name[:-4]
plt.xlabel("Testing Sample Index") # Removed unnecessary context from label
plt.ylabel("Remaining Useful Life (RUL)") # Removed unnecessary context from label
plt.title(f"Comparison of True and Predicted RUL for {dataset_name}") # Removed unnecessary context from title

# Legend and Grid
plt.legend()

# Save and Show
plt.savefig(f'figures/True_vs_PredictedRUL_simple_calibration_{file_name}.png')
plt.show()

# Calculate errors and mean predictions
errors = np.abs(y_pred - y)

# Scatter plot with enhancements
plt.figure(figsize=(8, 6))
sc = plt.scatter(margin_of_error, errors, c=errors, cmap='viridis', alpha=0.7, edgecolor='k')

# Add colorbar
cbar = plt.colorbar(sc)
cbar.set_label('Error Magnitude')

# Add grid and labels
plt.grid(alpha=0.3)
plt.xlabel('Uncertainty', fontsize=12)
plt.ylabel('Error', fontsize=12)
plt.title('Uncertainty vs. Error', fontsize=14)

# Polynomial fit (degree 2 or higher)
degree = 1  # try 3, 4, etc. for more complex trends
z = np.polyfit(margin_of_error, errors, degree)
p = np.poly1d(z)

# Sort for smooth plotting
sorted_idx = np.argsort(margin_of_error)
plt.plot(margin_of_error[sorted_idx], p(margin_of_error[sorted_idx]),
         color='red', linestyle='--', label=f'Polynomial Regression (deg {degree})')

# plt.plot(uncertainty, p(uncertainty), color='red', linestyle='--', label='Trend Line')

# Add legend
plt.legend()

# Show the plot
plt.tight_layout()
plt.savefig(f'figures\\UncertaintyVSError_simple_calibration_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()


# Compute statistics
correlation = np.corrcoef(margin_of_error, errors)[0, 1]
print(f"Correlation between uncertainty and error: {correlation:.2f}")

# Evaluate performance in high/low uncertainty regions
threshold = np.percentile(margin_of_error, 75)  # Upper 25% uncertainty
high_uncertainty_mask = margin_of_error > threshold
low_uncertainty_mask = margin_of_error <= threshold

high_uncertainty_error = errors[high_uncertainty_mask].mean()
low_uncertainty_error = errors[low_uncertainty_mask].mean()

print(f"Mean error in high uncertainty regions: {high_uncertainty_error:.2f}")
print(f"Mean error in low uncertainty regions: {low_uncertainty_error:.2f}")

Correlation between uncertainty and error: 0.22
Mean error in high uncertainty regions: 0.06
Mean error in low uncertainty regions: 0.07

def compute_coverage(y_true, y_pred, scaled_stds):
    """
    Computes the empirical coverage: the fraction of samples where
    the true value is inside the predicted confidence interval.
    
    Parameters:
    - y_true : np.array, ground truth values
    - y_pred : np.array, predicted means
    - scaled_stds : np.array, calibrated standard deviations (after applying k-scaling)

    Returns:
    - coverage : float, percentage of samples inside the confidence interval
    """
    lower_bounds = y_pred - scaled_stds
    upper_bounds = y_pred + scaled_stds

    # Check if the true value lies inside the interval
    inside = (y_true >= lower_bounds) & (y_true <= upper_bounds)

    # Calculate coverage as a percentage
    coverage = np.mean(inside) * 100.0
    return coverage


coverage = compute_coverage(y, y_pred, k_opt * uncertainty)
print(f"Ratio of coverage:{coverage}%")

Ratio of coverage:94.0%

# def calibrate_scaling_adaptive_by_mean(val_pred_means, val_stds, val_targets, num_bins=10, alpha=0.05):
#     """
#     Calibrates uncertainty scaling adaptively based on predicted mean value bins.
#     """
#     # Define bins based on predicted means, not errors.
#     mean_bins = np.linspace(np.min(val_pred_means), np.max(val_pred_means), num_bins + 1)
#     k_values = []

#     for i in range(num_bins):
#         bin_indices = np.where((val_pred_means >= mean_bins[i]) & (val_pred_means < mean_bins[i + 1]))[0]
#         if len(bin_indices) == 0:
#             k_values.append(1.0)  # Default if bin is empty.
#             continue

#         bin_means = val_pred_means[bin_indices]
#         bin_stds = val_stds[bin_indices]
#         bin_targets = val_targets[bin_indices]

#         def coverage_diff(k):
#             lower = bin_means - k * bin_stds
#             upper = bin_means + k * bin_stds
#             coverage = np.mean((bin_targets >= lower) & (bin_targets <= upper))
#             return abs((1 - alpha) - coverage)

#         result = optim.minimize_scalar(coverage_diff, bounds=(0.1, 5.0), method='bounded')
#         k_values.append(result.x)

#     return k_values, mean_bins


# def apply_adaptive_scaling_by_mean(pred_means, pred_stds, k_values, mean_bins):
#     """
#     Applies adaptive scaling based on predicted mean bins.
#     """
#     scaled_stds = np.zeros_like(pred_stds)
#     for i in range(len(k_values)):
#         bin_indices = np.where((pred_means >= mean_bins[i]) & (pred_means < mean_bins[i + 1]))[0]
#         scaled_stds[bin_indices] = pred_stds[bin_indices] * k_values[i]
#     return scaled_stds


# # === Example Usage ===

# # Fit k-values using validation data:
# k_values, mean_bins = calibrate_scaling_adaptive_by_mean(val_pred_means, val_pred_stds, val_targets, num_bins=5, alpha=0.05)

# # Apply learned k-values to a new (or same) dataset:
# scaled_stds = apply_adaptive_scaling_by_mean(val_pred_means, val_pred_stds, k_values, mean_bins)

# print("K values per predicted mean bin:", k_values)

# bin_midpoints = 0.5 * (mean_bins[:-1] + mean_bins[1:])
# plt.plot(bin_midpoints, k_values, marker='o')
# plt.xlabel('Predicted Mean (μ̂)')
# plt.ylabel('Learned Scaling Factor (k)')
# plt.title('Adaptive Scaling Factor per Predicted Mean Bin')
# plt.grid(True, linestyle='--', alpha=0.6)
# plt.show()

# plt.figure(figsize=(8, 6)) # Adjust figure size for publication

# plt.ylim(0, MAX_RUL+1)
# # True Values (more prominent line)
# plt.plot(range(len(y)), y * MAX_RUL, label="True RUL", c="lightseagreen", linewidth=2)

# # Boolean mask: True if point is inside the confidence interval
# inside_interval = (y >= y_pred - scaled_uncertaintity) & (y <= y_pred + scaled_uncertaintity)

# # Assign colors: green if inside, red if outside
# colors_interval = np.where(inside_interval, 'mediumseagreen', 'crimson')

# # Predicted Values (smaller scatter points)
# plt.scatter(range(len(y_pred)), y_pred * MAX_RUL, label="Predicted RUL", c=colors_interval, edgecolor="black", alpha=0.7, zorder=3, s=15) # reduced scatter point size to 15

# # Vertical Lines (improved clarity)
# for i in range(len(y)):
#     plt.plot([i, i], [y_pred[i]*MAX_RUL, y[i]*MAX_RUL], c="gray", linestyle="--", linewidth=0.8, alpha=0.7)

# # Confidence Interval (adjusted opacity, more descriptive label)
# plt.fill_between(
#     range(len(y_pred)),
#     (y_pred - scaled_uncertaintity) * MAX_RUL,
#     np.minimum(y_pred + scaled_uncertaintity, 1) * MAX_RUL,
#     color="salmon",
#     alpha=0.3,  # Reduced opacity
#     label=f"{int(confidence * 100)}% Prediction Interval"  # More accurate label
# )

# # Labels and Title (improved clarity)
# dataset_name = file_name[:-4]
# plt.xlabel("Testing Sample Index") # Removed unnecessary context from label
# plt.ylabel("Remaining Useful Life (RUL)") # Removed unnecessary context from label
# plt.title(f"Comparison of True and Predicted RUL for {dataset_name}") # Removed unnecessary context from title

# # Legend and Grid
# plt.legend()

# # Save and Show
# plt.savefig(f'figures/True_vs_PredictedRUL_simple_calibration_with_bins_{file_name}.png')
# plt.show()

# # Scatter plot with enhancements
# plt.figure(figsize=(8, 6))
# sc = plt.scatter(scaled_uncertaintity, test_errors, c=test_errors, cmap='viridis', alpha=0.7, edgecolor='k')

# # Add colorbar
# cbar = plt.colorbar(sc)
# cbar.set_label('Error Magnitude')

# # Add grid and labels
# plt.grid(alpha=0.3)
# plt.xlabel('Uncertainty', fontsize=12)
# plt.ylabel('Error', fontsize=12)
# plt.title('Uncertainty vs. Error', fontsize=14)

# # Polynomial fit (degree 2 or higher)
# degree = 1  # try 3, 4, etc. for more complex trends
# z = np.polyfit(scaled_uncertaintity, test_errors, degree)
# p = np.poly1d(z)

# # Sort for smooth plotting
# # sorted_idx = np.argsort(scaled_uncertaintity)
# plt.plot(scaled_uncertaintity, p(scaled_uncertaintity),
#           color='red', linestyle='--', label=f'Polynomial Regression (deg {degree})')

# # plt.plot(uncertainty, p(uncertainty), color='red', linestyle='--', label='Trend Line')

# # Add legend
# plt.legend()

# # Show the plot
# plt.tight_layout()
# plt.savefig(f'figures\\UncertaintyVSError_simple_calibration_with_bins_{file_name}.png', dpi=300, bbox_inches='tight')
# plt.show()


# # Compute statistics
# correlation = np.corrcoef(scaled_uncertaintity, test_errors)[0, 1]
# print(f"Correlation between uncertainty and error: {correlation:.2f}")

# # Evaluate performance in high/low uncertainty regions
# threshold = np.percentile(scaled_uncertaintity, 75)  # Upper 25% uncertainty
# high_uncertainty_mask = scaled_uncertaintity > threshold
# low_uncertainty_mask = scaled_uncertaintity <= threshold

# high_uncertainty_error = test_errors[high_uncertainty_mask].mean()
# low_uncertainty_error = test_errors[low_uncertainty_mask].mean()

# print(f"Mean error in high uncertainty regions: {high_uncertainty_error:.2f}")
# print(f"Mean error in low uncertainty regions: {low_uncertainty_error:.2f}")

from scipy import optimize as optim
from sklearn.cluster import KMeans
from sklearn.mixture import GaussianMixture
from sklearn.cluster import DBSCAN

X = np.vstack([val_pred_means, val_pred_stds]).T

# Fit Gaussian Mixture Model
num_components = 5
gmm = GaussianMixture(n_components=num_components, random_state=42).fit(X)
cluster_ids = gmm.predict(X)

# Plot clusters
plt.figure(figsize=(8, 6))
scatter = plt.scatter(val_pred_means, val_pred_stds, c=cluster_ids, cmap='tab10', s=40, edgecolor='k', alpha=0.7)

# Plot Gaussian ellipses
def plot_gmm_ellipses(gmm, ax):
    for n in range(gmm.n_components):
        mean = gmm.means_[n]
        cov = gmm.covariances_[n]

        if cov.shape == (2, 2):  # Full covariance matrix
            eigvals, eigvecs = np.linalg.eigh(cov)
            angle = np.degrees(np.arctan2(*eigvecs[:, 1][::-1]))
            width, height = 2 * np.sqrt(eigvals)  # 1-sigma contour
        else:  # Diagonal or spherical covariance
            width, height = 2 * np.sqrt(cov.ravel())
            angle = 0

        ellipse = Ellipse(xy=mean, width=width, height=height, angle=angle,
                          edgecolor='red', facecolor='none', linestyle='--', lw=2)
        ax.add_patch(ellipse)

ax = plt.gca()
plot_gmm_ellipses(gmm, ax)

plt.xlabel("Predicted Mean")
plt.ylabel("Predicted Std")
plt.title("GMM Clustering of Predictions")
plt.legend(*scatter.legend_elements(), title="Cluster")
plt.grid(True)
plt.show()

def calibrate_scaling_by_clustering(val_pred_means, val_stds, val_targets, num_clusters=3, alpha=0.05):
    """
    Clusters predictions by (mean, std) and calibrates a scaling factor k for each cluster.
    """
    X = np.vstack([val_pred_means, val_stds]).T
    kmeans = KMeans(n_clusters=num_clusters, random_state=42).fit(X)
    cluster_ids = kmeans.predict(X)

    k_values = []

    for cluster in range(num_clusters):
        bin_indices = np.where(cluster_ids == cluster)[0]
        if len(bin_indices) == 0:
            k_values.append(1.0)  # Default scaling if no samples fall here
            continue

        bin_means = val_pred_means[bin_indices]
        bin_stds = val_stds[bin_indices]
        bin_targets = val_targets[bin_indices]

        def coverage_diff(k):
            lower = bin_means - k * bin_stds
            upper = bin_means + k * bin_stds
            coverage = np.mean((bin_targets >= lower) & (bin_targets <= upper))
            return abs((1 - alpha) - coverage)

        result = optim.minimize_scalar(coverage_diff, bounds=(0.1, 5.0), method='bounded')
        k_values.append(result.x)

    return kmeans, k_values

def calibrate_scaling_by_gmm(val_pred_means, val_stds, val_targets, num_components=5, alpha=0.05):
    """
    Calibrates scaling factors k using Gaussian Mixture Model clustering on predicted mean and std.
    """
    X = np.vstack([val_pred_means, val_stds]).T
    gmm = GaussianMixture(n_components=num_components, random_state=42).fit(X)
    cluster_ids = gmm.predict(X)

    k_values = []
    for cluster in range(num_components):
        bin_indices = np.where(cluster_ids == cluster)[0]
        if len(bin_indices) == 0:
            k_values.append(1.0)  # Default scaling for empty clusters
            continue

        bin_means = val_pred_means[bin_indices]
        bin_stds = val_stds[bin_indices]
        bin_targets = val_targets[bin_indices]

        def coverage_diff(k):
            lower = bin_means - k * bin_stds
            upper = bin_means + k * bin_stds
            coverage = np.mean((bin_targets >= lower) & (bin_targets <= upper))
            return abs((1 - alpha) - coverage)

        result = optim.minimize_scalar(coverage_diff, bounds=(0.1, 5.0), method='bounded')
        k_values.append(result.x)

    return gmm, k_values

def calibrate_scaling_by_dbscan(val_pred_means, val_stds, val_targets, eps=0.1, min_samples=10, alpha=0.05):
    """
    Calibrates scaling factors k using DBSCAN clustering on predicted mean and std.
    """
    X = np.vstack([val_pred_means, val_stds]).T
    dbscan = DBSCAN(eps=eps, min_samples=min_samples).fit(X)
    cluster_ids = dbscan.labels_  # -1 means outlier

    unique_clusters = np.unique(cluster_ids)
    k_values = {}

    for cluster in unique_clusters:
        bin_indices = np.where(cluster_ids == cluster)[0]
        if len(bin_indices) == 0:
            continue  # Shouldn't happen

        bin_means = val_pred_means[bin_indices]
        bin_stds = val_stds[bin_indices]
        bin_targets = val_targets[bin_indices]

        def coverage_diff(k):
            lower = bin_means - k * bin_stds
            upper = bin_means + k * bin_stds
            coverage = np.mean((bin_targets >= lower) & (bin_targets <= upper))
            return abs((1 - alpha) - coverage)

        result = optim.minimize_scalar(coverage_diff, bounds=(0.1, 5.0), method='bounded')
        k_values[cluster] = result.x

    return dbscan, k_values


def apply_scaling_by_clustering(pred_means, pred_stds, kmeans, k_values):
    """
    Applies cluster-based scaling using learned k values.
    """
    X = np.vstack([pred_means, pred_stds]).T
    cluster_ids = kmeans.predict(X)

    scaled_stds = np.zeros_like(pred_stds)
    for i, cluster_id in enumerate(cluster_ids):
        scaled_stds[i] = pred_stds[i] * k_values[cluster_id]

    return scaled_stds

def apply_scaling_by_gmm(pred_means, pred_stds, gmm, k_values):
    """
    Applies scaling using GMM-clustered k values.
    """
    X = np.vstack([pred_means, pred_stds]).T
    cluster_ids = gmm.predict(X)

    scaled_stds = np.zeros_like(pred_stds)
    for i, cluster_id in enumerate(cluster_ids):
        scaled_stds[i] = pred_stds[i] * k_values[cluster_id]

    return scaled_stds

def apply_scaling_by_dbscan(pred_means, pred_stds, dbscan, k_values):
    """
    Applies scaling using DBSCAN-clustered k values. Assigns k=1.0 to outliers.
    """
    X = np.vstack([pred_means, pred_stds]).T
    cluster_ids = dbscan.fit_predict(X)  # Predict via re-fit for DBSCAN.

    scaled_stds = np.zeros_like(pred_stds)
    for i, cluster_id in enumerate(cluster_ids):
        k = k_values.get(cluster_id, 1.0)  # Outliers get k=1.0
        scaled_stds[i] = pred_stds[i] * k

    return scaled_stds


# === Example Usage ===

# Fit k-values on validation data:
# kmeans_model, k_values = calibrate_scaling_by_clustering(val_pred_means, val_pred_stds, val_targets, num_clusters=3, alpha=0.05)
gmm_model, k_values = calibrate_scaling_by_gmm(val_pred_means, val_pred_stds, val_targets, num_components=5, alpha=0.05)
# dbscan_model, k_values = calibrate_scaling_by_dbscan(val_pred_means, val_pred_stds, val_targets, eps=0.1, min_samples=10, alpha=0.05)

# Apply to test data:
# scaled_stds = apply_scaling_by_clustering(y_pred, uncertainty, kmeans_model, k_values)
scaled_stds =  apply_scaling_by_gmm(y_pred, uncertainty, gmm_model, k_values)
# scaled_stds =  apply_scaling_by_dbscan(y_pred, uncertainty, dbscan_model, k_values)

print("Cluster-wise learned k values:", k_values)

Cluster-wise learned k values: [0.7750665203530188, 3.128369953115367, 3.8054780887703274, 1.8945177908330837, 3.203146703495102]

plt.figure(figsize=(8, 6)) # Adjust figure size for publication

plt.ylim(0, MAX_RUL+1)
# True Values (more prominent line)
plt.plot(range(len(y)), y * MAX_RUL, label="True RUL", c="lightseagreen", linewidth=2)

# Boolean mask: True if point is inside the confidence interval
inside_interval = (y >= y_pred - scaled_stds) & (y <= y_pred + scaled_stds)

# Assign colors: green if inside, red if outside
colors_interval = np.where(inside_interval, 'mediumseagreen', 'crimson')

# Predicted Values (smaller scatter points)
plt.scatter(range(len(y_pred)), y_pred * MAX_RUL, label="Predicted RUL", c=colors_interval, edgecolor="black", alpha=0.7, zorder=3, s=15) # reduced scatter point size to 15

# Vertical Lines (improved clarity)
for i in range(len(y)):
    plt.plot([i, i], [y_pred[i]*MAX_RUL, y[i]*MAX_RUL], c="gray", linestyle="--", linewidth=0.8, alpha=0.7)

# Confidence Interval (adjusted opacity, more descriptive label)
plt.fill_between(
    range(len(y_pred)),
    (y_pred - 0.95*scaled_stds) * MAX_RUL,
    np.minimum(y_pred + 0.95*scaled_stds, 1) * MAX_RUL,
    color="salmon",
    alpha=0.3,  # Reduced opacity
    label=f"{int(confidence * 100)}% Prediction Interval"  # More accurate label
)

# Labels and Title (improved clarity)
dataset_name = file_name[:-4]
plt.xlabel("Testing Sample Index") # Removed unnecessary context from label
plt.ylabel("Remaining Useful Life (RUL)") # Removed unnecessary context from label
plt.title(f"Comparison of True and Predicted RUL for {dataset_name}") # Removed unnecessary context from title

# Legend and Grid
plt.legend()

# Save and Show
plt.savefig(f'figures/True_vs_PredictedRUL_calibration_with_clustering_bins_{file_name}.png')
plt.show()

def compute_coverage(y_true, y_pred, scaled_stds):
    """
    Computes the empirical coverage: the fraction of samples where
    the true value is inside the predicted confidence interval.
    
    Parameters:
    - y_true : np.array, ground truth values
    - y_pred : np.array, predicted means
    - scaled_stds : np.array, calibrated standard deviations (after applying k-scaling)

    Returns:
    - coverage : float, percentage of samples inside the confidence interval
    """
    lower_bounds = y_pred - scaled_stds
    upper_bounds = y_pred + scaled_stds

    # Check if the true value lies inside the interval
    inside = (y_true >= lower_bounds) & (y_true <= upper_bounds)

    # Calculate coverage as a percentage
    coverage = np.mean(inside) * 100.0
    return coverage


coverage = compute_coverage(y, y_pred, scaled_stds)
print(f"Ratio of coverage:{coverage}%")

Ratio of coverage:95.0%

# Example usage:
test_errors = np.abs(y - y_pred)

# Scatter plot with enhancements
plt.figure(figsize=(8, 6))
sc = plt.scatter(scaled_stds, test_errors, c=test_errors, cmap='viridis', alpha=0.7, edgecolor='k')

# Add colorbar
cbar = plt.colorbar(sc)
cbar.set_label('Error Magnitude')

# Add grid and labels
plt.grid(alpha=0.3)
plt.xlabel('Uncertainty', fontsize=12)
plt.ylabel('Error', fontsize=12)
plt.title('Uncertainty vs. Error', fontsize=14)

# Polynomial fit (degree 2 or higher)
degree = 1  # try 3, 4, etc. for more complex trends
z = np.polyfit(scaled_stds, test_errors, degree)
p = np.poly1d(z)

# Sort for smooth plotting
# sorted_idx = np.argsort(scaled_uncertaintity)
plt.plot(scaled_stds, p(scaled_stds),
          color='red', linestyle='--', label=f'Polynomial Regression (deg {degree})')

# plt.plot(uncertainty, p(uncertainty), color='red', linestyle='--', label='Trend Line')

# Add legend
plt.legend()

# Show the plot
plt.tight_layout()
plt.savefig(f'figures\\UncertaintyVSError_calibration_with_clustering_bins_{file_name}.png', dpi=300, bbox_inches='tight')
plt.show()


# Compute statistics
correlation = np.corrcoef(scaled_stds, test_errors)[0, 1]
print(f"Correlation between uncertainty and error: {correlation:.2f}")

# Evaluate performance in high/low uncertainty regions
threshold = np.percentile(scaled_stds, 75)  # Upper 25% uncertainty
high_uncertainty_mask = scaled_stds > threshold
low_uncertainty_mask = scaled_stds <= threshold

high_uncertainty_error = test_errors[high_uncertainty_mask].mean()
low_uncertainty_error = test_errors[low_uncertainty_mask].mean()

print(f"Mean error in high uncertainty regions: {high_uncertainty_error:.2f}")
print(f"Mean error in low uncertainty regions: {low_uncertainty_error:.2f}")

Correlation between uncertainty and error: 0.40
Mean error in high uncertainty regions: 0.10
Mean error in low uncertainty regions: 0.05

from scipy.stats import norm

def compute_coverage_at_confidence(y_true, y_pred, scaled_stds, confidence_level):
    """
    Computes the empirical coverage for a given confidence level.
    """
    alpha = 1 - confidence_level
    z = norm.ppf(1 - alpha / 2)  # two-tailed z-score

    lower_bounds = y_pred - z * scaled_stds
    upper_bounds = y_pred + z * scaled_stds

    inside = (y_true >= lower_bounds) & (y_true <= upper_bounds)
    return np.mean(inside)


def compute_ece(conf_levels, emp_coverages):
    """
    Computes Expected Calibration Error (ECE) — L1 distance between confidence and empirical coverage.
    """
    return np.mean(np.abs(np.array(emp_coverages) - np.array(conf_levels)))


def plot_calibration_curve_and_ece(y_true, y_pred, scaled_stds, num_points=20):
    """
    Plots the calibration curve and computes ECE.
    
    y_true: ground truth values
    y_pred: predicted means
    scaled_stds: scaled std deviations (after calibration)
    num_points: number of confidence levels to check between 50% and 99%
    """
    confidence_levels = np.linspace(0.5, 0.99, num_points)
    empirical_coverages = []

    for c in confidence_levels:
        coverage = compute_coverage_at_confidence(y_true, y_pred, scaled_stds, c)
        empirical_coverages.append(coverage)

    # Compute ECE
    ece = compute_ece(confidence_levels, empirical_coverages)

    # Plot
    plt.figure(figsize=(7, 7))
    plt.plot(confidence_levels, empirical_coverages, marker='o', label='Model Calibration')
    plt.plot([0, 1], [0, 1], linestyle='--', color='gray', label='Perfect Calibration')

    plt.xlabel('Expected Confidence Level')
    plt.ylabel('Empirical Coverage')
    plt.title(f'Calibration Curve (ECE = {ece:.4f})')
    plt.legend()
    plt.grid(True)
    plt.tight_layout()
    plt.show()

    return ece

ece = plot_calibration_curve_and_ece(y, y_pred, scaled_stds)
print(f"Expected Calibration Error (ECE): {ece:.4f}")

Expected Calibration Error (ECE): 0.1905

# def calibrate_scaling_adaptive(val_pred_means, val_stds, val_targets, num_bins=10, alpha=0.05):
#     """
#     Calibrates uncertainty scaling adaptively based on error magnitude.
#     """
#     errors = np.abs(val_targets - val_pred_means)
#     error_bins = np.linspace(np.min(errors), np.max(errors), num_bins + 1)
#     k_values = []

#     for i in range(num_bins):
#         bin_indices = np.where((errors >= error_bins[i]) & (errors < error_bins[i + 1]))[0]
#         if len(bin_indices) == 0:
#             k_values.append(1.0) #Default value if bin is empty.
#             continue
#         bin_means = val_pred_means[bin_indices]
#         bin_stds = val_stds[bin_indices]
#         bin_targets = val_targets[bin_indices]

#         def coverage_diff(k):
#             lower = bin_means - k * bin_stds
#             upper = bin_means + k * bin_stds
#             coverage = np.mean((bin_targets >= lower) & (bin_targets <= upper))
#             return abs((1 - alpha) - coverage)

#         result = optim.minimize_scalar(coverage_diff, bounds=(0.1, 5.0), method='bounded')
#         k_values.append(result.x)

#     return k_values, error_bins

# def apply_adaptive_scaling(pred_means, pred_stds, errors, k_values, error_bins):
#     """
#     Applies adaptive scaling based on error magnitude.
#     """
#     scaled_stds = np.zeros_like(pred_stds)
#     for i in range(len(k_values)):
#         bin_indices = np.where((errors >= error_bins[i]) & (errors < error_bins[i + 1]))[0]
#         scaled_stds[bin_indices] = pred_stds[bin_indices] * k_values[i]
#     return scaled_stds

# # Example usage:
# k_values, error_bins = calibrate_scaling_adaptive(val_pred_means, val_pred_stds, val_targets, num_bins=10, alpha=0.05)
# val_errors = np.abs(val_targets - val_pred_means)
# scaled_val_stds = apply_adaptive_scaling(val_pred_means, val_pred_stds, errors, k_values, error_bins)

# print("K values per error bin:", k_values)

# # Example usage:
# test_errors = np.abs(y - y_pred)
# scaled_uncertaintity = apply_adaptive_scaling(y, uncertainty, test_errors, k_values, error_bins)

# plt.figure(figsize=(8, 6)) # Adjust figure size for publication

# plt.ylim(0, MAX_RUL+1)
# # True Values (more prominent line)
# plt.plot(range(len(y)), y * MAX_RUL, label="True RUL", c="lightseagreen", linewidth=2)

# # Boolean mask: True if point is inside the confidence interval
# inside_interval = (y >= y_pred - scaled_uncertaintity) & (y <= y_pred + scaled_uncertaintity)

# # Assign colors: green if inside, red if outside
# colors_interval = np.where(inside_interval, 'mediumseagreen', 'crimson')

# # Predicted Values (smaller scatter points)
# plt.scatter(range(len(y_pred)), y_pred * MAX_RUL, label="Predicted RUL", c=colors_interval, edgecolor="black", alpha=0.7, zorder=3, s=15) # reduced scatter point size to 15

# # Vertical Lines (improved clarity)
# for i in range(len(y)):
#     plt.plot([i, i], [y_pred[i]*MAX_RUL, y[i]*MAX_RUL], c="gray", linestyle="--", linewidth=0.8, alpha=0.7)

# # Confidence Interval (adjusted opacity, more descriptive label)
# plt.fill_between(
#     range(len(y_pred)),
#     (y_pred - scaled_uncertaintity) * MAX_RUL,
#     np.minimum(y_pred + scaled_uncertaintity, 1) * MAX_RUL,
#     color="salmon",
#     alpha=0.3,  # Reduced opacity
#     label=f"{int(confidence * 100)}% Prediction Interval"  # More accurate label
# )

# # Labels and Title (improved clarity)
# dataset_name = file_name[:-4]
# plt.xlabel("Testing Sample Index") # Removed unnecessary context from label
# plt.ylabel("Remaining Useful Life (RUL)") # Removed unnecessary context from label
# plt.title(f"Comparison of True and Predicted RUL for {dataset_name}") # Removed unnecessary context from title

# # Legend and Grid
# plt.legend()

# # Save and Show
# plt.savefig(f'figures/True_vs_PredictedRUL_simple_calibration_with_bins_{file_name}.png')
# plt.show()

# # Scatter plot with enhancements
# plt.figure(figsize=(8, 6))
# sc = plt.scatter(scaled_uncertaintity, test_errors, c=test_errors, cmap='viridis', alpha=0.7, edgecolor='k')

# # Add colorbar
# cbar = plt.colorbar(sc)
# cbar.set_label('Error Magnitude')

# # Add grid and labels
# plt.grid(alpha=0.3)
# plt.xlabel('Uncertainty', fontsize=12)
# plt.ylabel('Error', fontsize=12)
# plt.title('Uncertainty vs. Error', fontsize=14)

# # Polynomial fit (degree 2 or higher)
# degree = 1  # try 3, 4, etc. for more complex trends
# z = np.polyfit(scaled_uncertaintity, test_errors, degree)
# p = np.poly1d(z)

# # Sort for smooth plotting
# # sorted_idx = np.argsort(scaled_uncertaintity)
# plt.plot(scaled_uncertaintity, p(scaled_uncertaintity),
#           color='red', linestyle='--', label=f'Polynomial Regression (deg {degree})')

# # plt.plot(uncertainty, p(uncertainty), color='red', linestyle='--', label='Trend Line')

# # Add legend
# plt.legend()

# # Show the plot
# plt.tight_layout()
# plt.savefig(f'figures\\UncertaintyVSError_simple_calibration_with_bins_{file_name}.png', dpi=300, bbox_inches='tight')
# plt.show()


# # Compute statistics
# correlation = np.corrcoef(scaled_uncertaintity, test_errors)[0, 1]
# print(f"Correlation between uncertainty and error: {correlation:.2f}")

# # Evaluate performance in high/low uncertainty regions
# threshold = np.percentile(scaled_uncertaintity, 75)  # Upper 25% uncertainty
# high_uncertainty_mask = scaled_uncertaintity > threshold
# low_uncertainty_mask = scaled_uncertaintity <= threshold

# high_uncertainty_error = test_errors[high_uncertainty_mask].mean()
# low_uncertainty_error = test_errors[low_uncertainty_mask].mean()

# print(f"Mean error in high uncertainty regions: {high_uncertainty_error:.2f}")
# print(f"Mean error in low uncertainty regions: {low_uncertainty_error:.2f}")

# import scipy.stats as st
# import scipy.optimize as optim

# def calibrate_scaling_by_prediction(val_pred_means, val_stds, val_targets, alpha=0.05, func_type='linear'):
#     """
#     Calibrates uncertainty scaling using a function of the predicted mean (μ).
#     Learns parameters such that the prediction intervals cover (1 - alpha) of the targets.
#     """

#     def compute_coverage(k_values):
#         lower = val_pred_means - k_values * val_stds
#         upper = val_pred_means + k_values * val_stds
#         return np.mean((val_targets >= lower) & (val_targets <= upper))

#     def coverage_diff_linear(params):
#         a, b = params
#         k_values = np.clip(a * val_pred_means + b, 0.01, 10.0)  # prevent negative scaling
#         return abs((1 - alpha) - compute_coverage(k_values))

#     def coverage_diff_quadratic(params):
#         a, b, c = params
#         k_values = np.clip(a * val_pred_means**2 + b * val_pred_means + c, 0.01, 10.0)
#         return abs((1 - alpha) - compute_coverage(k_values))

#     if func_type == 'linear':
#         initial_params = [0.0, 1.0]  # Start from identity scaling
#         result = optim.minimize(
#             coverage_diff_linear,
#             initial_params,
#             method='L-BFGS-B',
#             bounds=[(None, None), (0.1, 5.0)]
#         )
#         optimal_params = result.x
#         print(f"Optimal parameters (linear): {optimal_params}")
#         return lambda pred: np.clip(optimal_params[0] * pred + optimal_params[1], 0.01, 10.0)

#     elif func_type == 'quadratic':
#         initial_params = [0.0, 0.0, 1.0]  # Start from no correction
#         result = optim.minimize(
#             coverage_diff_quadratic,
#             initial_params,
#             method='L-BFGS-B',
#             bounds=[(None, None), (None, None), (0.1, 5.0)]
#         )
#         optimal_params = result.x
#         print(f"Optimal parameters (quadratic): {optimal_params}")
#         return lambda pred: np.clip(optimal_params[0] * pred**2 + optimal_params[1] * pred + optimal_params[2], 0.01, 10.0)

#     else:
#         raise ValueError(f"Unknown func_type: {func_type}")


# def apply_prediction_based_scaling(pred_means, pred_stds, k_function):
#     """
#     Applies the learned scaling function at inference time.
#     """
#     k_values = k_function(pred_means)  # Vectorized, no loop needed
#     scaled_stds = pred_stds * k_values
#     return scaled_stds

# # === EXAMPLE USAGE ===

# # During validation: learn the scaling function
# k_function_pred = calibrate_scaling_by_prediction(
#     val_pred_means, val_stds=val_pred_stds, val_targets=val_targets,
#     alpha=0.05, func_type='quadratic'  #  'linear' or 'quadratic'
# )

# # During inference: apply the learned calibration function
# scaled_test_stds = apply_prediction_based_scaling(y_pred, uncertainty, k_function_pred)

# # Your prediction intervals are now:
# # lower_bound = y_pred - scaled_test_stds
# # upper_bound = y_pred + scaled_test_stds

# print("✅ Calibration function learned and applied.")

# plt.figure(figsize=(8, 6)) # Adjust figure size for publication

# # True Values (more prominent line)
# plt.plot(range(len(y)), y, label="True RUL", c="lightseagreen", linewidth=2)

# # Boolean mask: True if point is inside the confidence interval
# inside_interval = (y >= y_pred - scaled_uncertaintity) & (y <= y_pred + scaled_uncertaintity)

# # Assign colors: green if inside, red if outside
# colors_interval = np.where(inside_interval, 'mediumseagreen', 'crimson')

# # Predicted Values (smaller scatter points)
# plt.scatter(range(len(y_pred)), y_pred, label="Predicted RUL", c=colors_interval, edgecolor="black", alpha=0.7, zorder=3, s=15) # reduced scatter point size to 15

# # Vertical Lines (improved clarity)
# for i in range(len(y)):
#     plt.plot([i, i], [y_pred[i], y[i]], c="gray", linestyle="--", linewidth=0.8, alpha=0.7)

# # Confidence Interval (adjusted opacity, more descriptive label)
# plt.fill_between(
#     range(len(y_pred)),
#     y_pred - scaled_uncertaintity,
#     y_pred + scaled_uncertaintity,
#     color="salmon",
#     alpha=0.3,  # Reduced opacity
#     label=f"{int(confidence * 100)}% Prediction Interval"  # More accurate label
# )

# # Labels and Title (improved clarity)
# dataset_name = file_name[:-4]
# plt.xlabel("Testing Sample Index") # Removed unnecessary context from label
# plt.ylabel("Remaining Useful Life (RUL)") # Removed unnecessary context from label
# plt.title(f"Comparison of True and Predicted RUL for {dataset_name}") # Removed unnecessary context from title

# # Legend and Grid
# plt.legend()

# # Save and Show
# # plt.savefig(f'figures/True_vs_PredictedRUL_{file_name}.png')
# plt.show()

# # Scatter plot with enhancements
# plt.figure(figsize=(8, 6))
# sc = plt.scatter(scaled_uncertaintity, test_errors, c=test_errors, cmap='viridis', alpha=0.7, edgecolor='k')

# # Add colorbar
# cbar = plt.colorbar(sc)
# cbar.set_label('Error Magnitude')

# # Add grid and labels
# plt.grid(alpha=0.3)
# plt.xlabel('Uncertainty', fontsize=12)
# plt.ylabel('Error', fontsize=12)
# plt.title('Uncertainty vs. Error', fontsize=14)

# # Polynomial fit (degree 2 or higher)
# degree = 1  # try 3, 4, etc. for more complex trends
# z = np.polyfit(scaled_uncertaintity, test_errors, degree)
# p = np.poly1d(z)

# # Sort for smooth plotting
# # sorted_idx = np.argsort(scaled_uncertaintity)
# plt.plot(scaled_uncertaintity, p(scaled_uncertaintity),
#           color='red', linestyle='--', label=f'Polynomial Regression (deg {degree})')

# # plt.plot(uncertainty, p(uncertainty), color='red', linestyle='--', label='Trend Line')

# # Add legend
# plt.legend()

# # Show the plot
# plt.tight_layout()
# # plt.savefig(f'figures\\UncertaintyVSError_{file_name}.png', dpi=300, bbox_inches='tight')
# plt.show()


# # Compute statistics
# correlation = np.corrcoef(scaled_uncertaintity, test_errors)[0, 1]
# print(f"Correlation between uncertainty and error: {correlation:.2f}")

# # Evaluate performance in high/low uncertainty regions
# threshold = np.percentile(scaled_uncertaintity, 75)  # Upper 25% uncertainty
# high_uncertainty_mask = scaled_uncertaintity > threshold
# low_uncertainty_mask = scaled_uncertaintity <= threshold

# high_uncertainty_error = test_errors[high_uncertainty_mask].mean()
# low_uncertainty_error = test_errors[low_uncertainty_mask].mean()

# print(f"Mean error in high uncertainty regions: {high_uncertainty_error:.2f}")
# print(f"Mean error in low uncertainty regions: {low_uncertainty_error:.2f}")

# def plot_k_function(k_function, val_pred_means, title="Learned Calibration Function: k(μ)"):
#     """
#     Plots the learned scaling function k(μ) over the range of predicted means.
#     """
#     mu_min = np.min(val_pred_means) - 0.1 * abs(np.min(val_pred_means))
#     mu_max = np.max(val_pred_means) + 0.1 * abs(np.max(val_pred_means))
#     mu_range = np.linspace(mu_min, mu_max, 500)

#     k_values = k_function(mu_range)

#     plt.figure(figsize=(8, 5))
#     plt.plot(mu_range, k_values, label='$k(\\hat{\\mu})$', color='royalblue', linewidth=2)
#     plt.xlabel('Predicted Mean $\\hat{\\mu}$', fontsize=12)
#     plt.ylabel('Scaling Factor $k$', fontsize=12)
#     plt.title(title, fontsize=14)
#     plt.grid(True, linestyle='--', alpha=0.6)
#     plt.legend()
#     plt.show()

# # === Example usage ===

# # Assuming you already ran:
# # k_function_pred = calibrate_scaling_by_prediction(val_pred_means, val_pred_stds, val_targets, alpha=0.05, func_type='quadratic')

# plot_k_function(k_function_pred, val_pred_means)

# import shap

# X_list, y_list = [], []

# for batch in valloader:
#     inputs, targets = batch
#     inputs, targets = inputs.to(device).float(), targets.to(device).float()
#     X_list.append(inputs)
#     y_list.append(targets)

# X_np = torch.cat(X_list).cpu().numpy()
# y_np = torch.cat(y_list).cpu().numpy()

# model.eval()

# def model_predict(x_numpy_flat):
#     # Reshape back to original shape before passing to model
#     x_numpy = x_numpy_flat.reshape(-1, 15, 17)
#     x_tensor = torch.from_numpy(x_numpy).float()
#     with torch.no_grad():
#         output = model(x_tensor)
#     return output.view(-1).cpu().numpy()

# # Background subset
# background = X_np[np.random.choice(X_np.shape[0], 100, replace=False)]

# X_flat = X_np.reshape(X_np.shape[0], -1)  # shape: (n_samples, 255)
# background_flat = background.reshape(background.shape[0], -1)  # same here


# # Create explainer
# explainer =  shap.DeepExplainer(model_predict, background_flat)

# # Compute SHAP values for a few samples
# shap_values = explainer.shap_values(X_flat[:10])

# # Plot
# shap.summary_plot(shap_values, X_flat[:10])

# from sklearn.tree import DecisionTreeRegressor
# from sklearn.metrics import mean_squared_error

# X_list, y_list, ypred_list =  [], [], []

# for batch in valloader:
#     inputs, targets = batch
#     inputs, targets = inputs.to(device).float(), targets.to(device).float()
#     X_list.append(inputs)
#     y_list.append(targets)
#     ypred_list.append(model(inputs).squeeze())

# X_np = torch.cat(X_list).cpu().numpy()
# y_np = torch.cat(y_list).cpu().numpy()
# ypred_np = torch.cat(ypred_list).detach().cpu().numpy()
# seq, fea = X_np.shape[1], X_np.shape[2]

# X_np = X_np.reshape(X_np.shape[0], -1)

# print(X_np.shape)
# print(y_np.shape)
# print(ypred_np.shape)


# surrogate_model = DecisionTreeRegressor(random_state=42)
# surrogate_model.fit(X_np, ypred_np)
# ypred_surrogate = surrogate_model.predict(X_np)

# mse = mean_squared_error(ypred_surrogate, ypred_np)
# print(f"Mean Squared Error of Surrogate Model: {mse:.4f}")
# feature_importance = surrogate_model.feature_importances_

# # Reshape feature importance back to (time, features)
# importance_matrix = feature_importance.reshape(seq, fea)  # (seq_len, num_features)
# log_importance = np.log10(importance_matrix + 1e-6)

# # Plot heatmap
# plt.figure(figsize=(8, 5))
# sns.heatmap(log_importance.T, cmap="viridis", annot=True, fmt=".2f", cbar=True,
#             xticklabels=[f"T{i+1}" for i in range(30)],
#             yticklabels=[f"F{j+1}" for j in range(4)])

# plt.title("Log-Scaled Feature Importance (Time × Feature)", fontsize=13, weight='bold')
# plt.xlabel("Time Step")
# plt.ylabel("Input Feature")
# plt.tight_layout()
# plt.show()

	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27
0	1	1	-0.0005	0.0004	100.0	518.67	642.36	1583.23	1396.84	14.62	21.61	553.97	2387.96	9062.17	1.3	47.30	522.31	2388.01	8145.32	8.4246	0.03	391	2388	100.0	39.11	23.3537	NaN	NaN
1	1	2	0.0008	-0.0003	100.0	518.67	642.50	1584.69	1396.89	14.62	21.61	554.55	2388.00	9061.78	1.3	47.23	522.42	2388.03	8152.85	8.4403	0.03	392	2388	100.0	38.99	23.4491	NaN	NaN
2	1	3	-0.0014	-0.0002	100.0	518.67	642.18	1582.35	1405.61	14.62	21.61	554.43	2388.03	9070.23	1.3	47.22	522.03	2388.00	8150.17	8.3901	0.03	391	2388	100.0	38.85	23.3669	NaN	NaN
3	1	4	-0.0020	0.0001	100.0	518.67	642.92	1585.61	1392.27	14.62	21.61	555.21	2388.00	9064.57	1.3	47.24	522.49	2388.08	8146.56	8.3878	0.03	392	2388	100.0	38.96	23.2951	NaN	NaN
4	1	5	0.0016	0.0000	100.0	518.67	641.68	1588.63	1397.65	14.62	21.61	554.74	2388.04	9076.14	1.3	47.15	522.58	2388.03	8147.80	8.3869	0.03	392	2388	100.0	39.14	23.4583	NaN	NaN

	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27
0	1	1	-0.0017	-0.0004	100.0	518.67	641.94	1581.93	1396.93	14.62	21.58	554.56	2387.93	9048.65	1.3	47.09	521.89	2387.94	8133.48	8.3760	0.03	391	2388	100.0	39.07	23.4468	NaN	NaN
1	1	2	0.0006	-0.0002	100.0	518.67	642.02	1584.86	1398.90	14.62	21.58	554.10	2387.94	9046.53	1.3	47.08	521.85	2388.01	8137.44	8.4062	0.03	391	2388	100.0	39.04	23.4807	NaN	NaN
2	1	3	0.0014	-0.0003	100.0	518.67	641.68	1581.78	1391.92	14.62	21.58	554.41	2387.97	9054.92	1.3	47.15	522.10	2387.94	8138.25	8.3553	0.03	391	2388	100.0	39.10	23.4244	NaN	NaN
3	1	4	0.0027	0.0001	100.0	518.67	642.20	1584.53	1395.34	14.62	21.59	554.58	2387.94	9055.04	1.3	47.26	522.45	2387.96	8137.07	8.3709	0.03	392	2388	100.0	38.97	23.4782	NaN	NaN
4	1	5	-0.0001	0.0001	100.0	518.67	642.46	1589.03	1395.86	14.62	21.58	554.16	2388.01	9048.59	1.3	46.94	521.91	2387.97	8134.20	8.4146	0.03	391	2388	100.0	39.09	23.3950	NaN	NaN

	unit	time	os1	os2	os3	s1	s2	s3	s4	s5	s6	s7	s8	s9	s10	s11	s12	s13	s14	s15	s16	s17	s18	s19	s20	s21
0	1	1	-0.0005	0.0004	100.0	518.67	642.36	1583.23	1396.84	14.62	21.61	553.97	2387.96	9062.17	1.3	47.30	522.31	2388.01	8145.32	8.4246	0.03	391	2388	100.0	39.11	23.3537
1	1	2	0.0008	-0.0003	100.0	518.67	642.50	1584.69	1396.89	14.62	21.61	554.55	2388.00	9061.78	1.3	47.23	522.42	2388.03	8152.85	8.4403	0.03	392	2388	100.0	38.99	23.4491
2	1	3	-0.0014	-0.0002	100.0	518.67	642.18	1582.35	1405.61	14.62	21.61	554.43	2388.03	9070.23	1.3	47.22	522.03	2388.00	8150.17	8.3901	0.03	391	2388	100.0	38.85	23.3669
3	1	4	-0.0020	0.0001	100.0	518.67	642.92	1585.61	1392.27	14.62	21.61	555.21	2388.00	9064.57	1.3	47.24	522.49	2388.08	8146.56	8.3878	0.03	392	2388	100.0	38.96	23.2951
4	1	5	0.0016	0.0000	100.0	518.67	641.68	1588.63	1397.65	14.62	21.61	554.74	2388.04	9076.14	1.3	47.15	522.58	2388.03	8147.80	8.3869	0.03	392	2388	100.0	39.14	23.4583

	unit	time	os1	os2	os3	s1	s2	s3	s4	s5	s6	s7	s8	s9	s10	s11	s12	s13	s14	s15	s16	s17	s18	s19	s20	s21
0	1	1	-0.0017	-0.0004	100.0	518.67	641.94	1581.93	1396.93	14.62	21.58	554.56	2387.93	9048.65	1.3	47.09	521.89	2387.94	8133.48	8.3760	0.03	391	2388	100.0	39.07	23.4468
1	1	2	0.0006	-0.0002	100.0	518.67	642.02	1584.86	1398.90	14.62	21.58	554.10	2387.94	9046.53	1.3	47.08	521.85	2388.01	8137.44	8.4062	0.03	391	2388	100.0	39.04	23.4807
2	1	3	0.0014	-0.0003	100.0	518.67	641.68	1581.78	1391.92	14.62	21.58	554.41	2387.97	9054.92	1.3	47.15	522.10	2387.94	8138.25	8.3553	0.03	391	2388	100.0	39.10	23.4244
3	1	4	0.0027	0.0001	100.0	518.67	642.20	1584.53	1395.34	14.62	21.59	554.58	2387.94	9055.04	1.3	47.26	522.45	2387.96	8137.07	8.3709	0.03	392	2388	100.0	38.97	23.4782
4	1	5	-0.0001	0.0001	100.0	518.67	642.46	1589.03	1395.86	14.62	21.58	554.16	2388.01	9048.59	1.3	46.94	521.91	2387.97	8134.20	8.4146	0.03	391	2388	100.0	39.09	23.3950

	unit	time	os1	os2	os3	s1	s2	s3	s4	s5	s6	s7	s8	s9	s10	s11	s12	s13	s14	s15	s16	s17	s18	s19	s20	s21	rul
0	1	1	-0.0005	0.0004	100.0	518.67	642.36	1583.23	1396.84	14.62	21.61	553.97	2387.96	9062.17	1.3	47.30	522.31	2388.01	8145.32	8.4246	0.03	391	2388	100.0	39.11	23.3537	125
1	1	2	0.0008	-0.0003	100.0	518.67	642.50	1584.69	1396.89	14.62	21.61	554.55	2388.00	9061.78	1.3	47.23	522.42	2388.03	8152.85	8.4403	0.03	392	2388	100.0	38.99	23.4491	125
2	1	3	-0.0014	-0.0002	100.0	518.67	642.18	1582.35	1405.61	14.62	21.61	554.43	2388.03	9070.23	1.3	47.22	522.03	2388.00	8150.17	8.3901	0.03	391	2388	100.0	38.85	23.3669	125
3	1	4	-0.0020	0.0001	100.0	518.67	642.92	1585.61	1392.27	14.62	21.61	555.21	2388.00	9064.57	1.3	47.24	522.49	2388.08	8146.56	8.3878	0.03	392	2388	100.0	38.96	23.2951	125
4	1	5	0.0016	0.0000	100.0	518.67	641.68	1588.63	1397.65	14.62	21.61	554.74	2388.04	9076.14	1.3	47.15	522.58	2388.03	8147.80	8.3869	0.03	392	2388	100.0	39.14	23.4583	125

2. Model Building and Training¶

3. Prediction on Test Set¶

4. Prediction on Test Set Without Dropout in the Inference¶

5. Methods for Weighting Predictions by Uncertainty¶

4. Temperature Scaling - Calibration on Uncertaintity¶

Intepredability Of the Model¶

	unit	time	os1	os2	s2	s3	s4	s5	s6	s7	s8	s9	s10	s11	s12	s13	s14	s15	s16	s17	s20	s21	rul
0	1	1	-0.217019	1.338555	-0.187098	-0.712023	-0.780832	-3.552714e-09	0.78153	-0.341487	-0.704766	-0.097136	-0.35337	-0.385592	-0.227589	-0.389847	0.067685	0.469724	-1.734723e-11	-0.889345	0.488007	-0.263504	125
1	1	2	0.375359	-1.033981	0.080572	-0.497646	-0.775716	-3.552714e-09	0.78153	-0.172752	-0.452059	-0.116655	-0.35337	-0.618867	-0.193798	-0.263362	0.523935	0.729174	-1.734723e-11	-0.321634	0.005819	0.375757	125
2	1	3	-0.627127	-0.695047	-0.531245	-0.841237	0.116522	-3.552714e-09	0.78153	-0.207663	-0.262529	0.306261	-0.35337	-0.652192	-0.313602	-0.453089	0.361551	-0.100405	-1.734723e-11	-0.889345	-0.556733	-0.175053	125
3	1	4	-0.900532	0.321754	0.883582	-0.362558	-1.248438	-3.552714e-09	0.78153	0.019257	-0.452059	0.022982	-0.35337	-0.585542	-0.172294	0.052850	0.142818	-0.138414	-1.734723e-11	-0.321634	-0.114728	-0.656174	125
4	1	5	0.739899	-0.017180	-1.487209	0.080880	-0.697952	-3.552714e-09	0.78153	-0.117477	-0.199352	0.602053	-0.35337	-0.885467	-0.144647	-0.263362	0.217951	-0.153287	-1.734723e-11	-0.321634	0.608554	0.437404	125

	unit	time	os1	os2	s2	s3	s4	s5	s6	s7	s8	s9	s10	s11	s12	s13	s14	s15	s16	s17	s20	s21	rul
0	1	1	-0.763830	-1.372915	-0.990107	-0.902907	-0.771623	-3.552714e-09	-0.874373	-0.169843	-0.894297	-0.773803	-0.35337	-1.085417	-0.356608	-0.832544	-0.649712	-0.333415	-1.734723e-11	-0.889345	0.327278	0.360345	125
1	1	2	0.284224	-0.695047	-0.837153	-0.472684	-0.570051	-3.552714e-09	-0.874373	-0.303667	-0.831120	-0.879907	-0.35337	-1.118742	-0.368896	-0.389847	-0.409771	0.165655	-1.734723e-11	-0.889345	0.206731	0.587504	125
2	1	3	0.648764	-1.033981	-1.487209	-0.924932	-1.284251	-3.552714e-09	-0.874373	-0.213481	-0.641590	-0.459994	-0.35337	-0.885467	-0.292098	-0.832544	-0.360693	-0.675492	-1.734723e-11	-0.889345	0.447825	0.210246	125
3	1	4	1.241142	0.321754	-0.493006	-0.521139	-0.934313	-3.552714e-09	-0.322405	-0.164024	-0.831120	-0.453988	-0.35337	-0.518892	-0.184582	-0.706059	-0.432190	-0.417694	-1.734723e-11	-0.321634	-0.074545	0.570751	125
4	1	5	-0.034749	0.321754	0.004095	0.139613	-0.881107	-3.552714e-09	-0.874373	-0.286212	-0.388882	-0.776806	-0.35337	-1.585292	-0.350465	-0.642816	-0.606086	0.304469	-1.734723e-11	-0.889345	0.407642	0.013241	125

	s5	s1	s16	s18	s19	os3	s10
count	2.472000e+04	24720.0	2.472000e+04	24720.0	24720.0	24720.0	2.472000e+04
mean	-3.552714e-09	0.0	-1.734723e-11	0.0	0.0	0.0	-8.342324e-14
std	4.135987e-25	0.0	3.231240e-27	0.0	0.0	0.0	9.997131e-01
min	-3.552714e-09	0.0	-1.734723e-11	0.0	0.0	0.0	-3.222112e+00
25%	-3.552714e-09	0.0	-1.734723e-11	0.0	0.0	0.0	-3.533705e-01
50%	-3.552714e-09	0.0	-1.734723e-11	0.0	0.0	0.0	-3.533705e-01
75%	-3.552714e-09	0.0	-1.734723e-11	0.0	0.0	0.0	-3.533705e-01
max	-3.552714e-09	0.0	-1.734723e-11	0.0	0.0	0.0	5.384113e+00

	unit	time	os1	os2	s2	s3	s4	s6	s7	s8	s9	s11	s12	s13	s14	s15	s17	s20	s21	rul
0	1	-1.396881	-0.217019	1.338555	-0.187098	-0.712023	-0.780832	0.78153	-0.341487	-0.704766	-0.097136	-0.385592	-0.227589	-0.389847	0.067685	0.469724	-0.889345	0.488007	-0.263504	1.000
1	1	-1.386765	0.082162	0.140305	-0.051911	-0.603752	-0.778248	0.78153	-0.256267	-0.577137	-0.106994	-0.503408	-0.210522	-0.325965	0.298114	0.600759	-0.602622	0.244478	0.059355	1.000
2	1	-1.376648	-0.159060	-0.143790	-0.214927	-0.684518	-0.473946	0.78153	-0.239737	-0.470142	0.033550	-0.554008	-0.245579	-0.369199	0.319689	0.362300	-0.700134	-0.028006	-0.020365	1.000
3	1	-1.366531	-0.350083	-0.023854	0.068078	-0.601573	-0.673476	0.78153	-0.173014	-0.465483	0.030827	-0.562132	-0.226699	-0.260468	0.274122	0.233303	-0.602622	-0.050348	-0.184166	1.000
4	1	-1.356414	-0.123190	-0.022464	-0.255673	-0.459512	-0.678571	0.78153	-0.161453	-0.410085	0.149734	-0.629438	-0.209619	-0.261070	0.262429	0.152830	-0.544131	0.086810	-0.054779	1.000
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
24715	100	0.090270	0.142461	-0.166179	1.153957	0.975020	1.250410	0.78153	-0.749628	0.654728	-0.351183	1.280620	-0.719768	0.632469	-0.580150	1.302012	1.086300	-1.285265	-1.267812	0.032
24716	100	0.100387	0.172284	-0.184435	1.163151	0.980149	1.277205	0.78153	-0.753956	0.662013	-0.343280	1.303755	-0.724923	0.646884	-0.567745	1.316765	1.092510	-1.299525	-1.300125	0.024
24717	100	0.110504	0.153569	-0.152430	1.194240	1.010314	1.308290	0.78153	-0.761611	0.671792	-0.336761	1.329882	-0.732999	0.655665	-0.566323	1.332893	1.110514	-1.322759	-1.318544	0.016
24718	100	0.120621	0.128570	-0.121128	1.228660	1.018329	1.328294	0.78153	-0.772824	0.682682	-0.331678	1.346339	-0.737222	0.666908	-0.558203	1.346065	1.116203	-1.352230	-1.335558	0.008
24719	100	0.130737	0.126100	-0.097623	1.250293	1.030726	1.368224	0.78153	-0.782143	0.693334	-0.322193	1.367327	-0.744766	0.675251	-0.555674	1.359017	1.133673	-1.359986	-1.350879	0.000

	0
0	44
1	51
2	27
3	120
4	101

	unit	time	os1	os2	s2	s3	s4	s6	s7	s8	s9	s11	s12	s13	s14	s15	s17	s20	s21	rul
0	1	1	-0.763830	-1.372915	-0.990107	-0.902907	-0.771623	-0.874373	-0.169843	-0.894297	-0.773803	-1.085417	-0.356608	-0.832544	-0.649712	-0.333415	-0.889345	0.327278	0.360345	125
1	1	2	-0.234510	-1.030558	-0.912858	-0.685623	-0.669819	-0.874373	-0.237431	-0.862389	-0.827391	-1.102248	-0.362814	-0.608959	-0.528530	-0.081359	-0.889345	0.266396	0.475071	125
2	1	3	0.065883	-1.031722	-1.108189	-0.767009	-0.878781	-0.874373	-0.229286	-0.787298	-0.702443	-1.028523	-0.338765	-0.684998	-0.471450	-0.283418	-0.889345	0.328098	0.385007	125
3	1	4	0.368660	-0.683031	-0.949702	-0.703667	-0.893088	-0.732172	-0.212473	-0.798587	-0.638434	-0.897229	-0.299043	-0.690424	-0.461336	-0.318011	-0.743088	0.224366	0.432860	125
4	1	5	0.284686	-0.473873	-0.751158	-0.528128	-0.890594	-0.761773	-0.227822	-0.713303	-0.667238	-1.040457	-0.309747	-0.680514	-0.491467	-0.188435	-0.773533	0.262517	0.345511	125